Journal of Computer Science December 2013

International Journal of Computer Science

& Information Security

© IJCSIS PUBLICATION 2013

IJCSIS Vol. 11 No. 12, December 2013 ISSN 1947-5500

IJCSIS

ISSN (online): 1947-5500

Please consider to contribute to and/or forward to the appropriate groups the following opportunity to submit and publish original scientific results. CALL FOR PAPERS International Journal of Computer Science and Information Security (IJCSIS) January-December 2014 Issues The topics suggested by this issue can be discussed in term of concepts, surveys, state of the art, research, standards, implementations, running experiments, applications, and industrial case studies. Authors are invited to submit complete unpublished papers, which are not under review in any other conference or journal in the following, but not limited to, topic areas. See authors guide for manuscript preparation and submission guidelines. Indexed by Google Scholar, DBLP, CiteSeerX, Directory for Open Access Journal (DOAJ), Bielefeld Academic Search Engine (BASE), SCIRUS, Scopus Database, Cornell University Library, ScientificCommons, ProQuest, EBSCO and more.

Deadline: see web site Notification: see web siteRevision: see web sitePublication: see web site

For more topics, please see web site https://sites.google.com/site/ijcsis/

For more information, please visit the journal website (https://sites.google.com/site/ijcsis/)

Context-aware systems Networking technologies Security in network, systems, and applications Evolutionary computation Industrial systems Evolutionary computation Autonomic and autonomous systems Bio-technologies Knowledge data systems Mobile and distance education Intelligent techniques, logics and systems Knowledge processing Information technologies Internet and web technologies Digital information processing Cognitive science and knowledge

Agent-based systems Mobility and multimedia systems Systems performance Networking and telecommunications Software development and deployment Knowledge virtualization Systems and networks on the chip Knowledge for global defense Information Systems [IS] IPv6 Today - Technology and deployment Modeling Software Engineering Optimization Complexity Natural Language Processing Speech Synthesis Data Mining

Editorial Message from Managing Editor

International Journal of Computer Science and Information Security (IJCSIS – established since May 2009), is a global venue to promote research and development results of high significance in the theory, design, implementation, analysis, and application of computing and security. As a scholarly open access peer-reviewed international journal, the main objective is to provide the academic community and industry a forum for dissemination of original research related to Computer Science and Security. High caliber authors regularly contribute to this journal by submitting articles that illustrate research results, projects, surveying works and industrial experiences relevant to latest advances in the Computer Science & Information Security.

IJCSIS archives all publications in major academic/scientific databases; abstracting/indexing, editorial board and other important information are available online on homepage. Indexed by the following International agencies and institutions: Google Scholar, Bielefeld Academic Search Engine (BASE), CiteSeerX, SCIRUS, Cornell’s University Library EI, Scopus, DBLP, DOI, ProQuest, EBSCO. Google Scholar reported increased in number cited papers published in IJCSIS. IJCSIS supports the Open Access policy of distribution of published manuscripts, ensuring "free availability on the public Internet, permitting any users to read, download, copy, distribute, print, search, or link to the full texts of [published] articles". IJCSIS editorial board ensures a rigorous peer-reviewing process and consisting of international experts. IJCSIS solicits your contribution with your research papers. IJCSIS is grateful for all the insights and advice from authors & reviewers. We look forward to your collaboration. Get in touch with us. For further questions please do not hesitate to contact us at [email protected]. A complete list of journals can be found at: http://sites.google.com/site/ijcsis/

IJCSIS Vol. 11, No. 12, December 2013 Edition

ISSN 1947-5500 © IJCSIS, USA.

Journal Indexed by (among others):

IJCSIS EDITORIAL BOARD Dr. Yong Li School of Electronic and Information Engineering, Beijing Jiaotong University, P. R. China Prof. Hamid Reza Naji Department of Computer Enigneering, Shahid Beheshti University, Tehran, Iran Dr. Sanjay Jasola Professor and Dean, School of Information and Communication Technology, Gautam Buddha University Dr Riktesh Srivastava Assistant Professor, Information Systems, Skyline University College, University City of Sharjah, Sharjah, PO 1797, UAE Dr. Siddhivinayak Kulkarni University of Ballarat, Ballarat, Victoria, Australia Professor (Dr) Mokhtar Beldjehem Sainte-Anne University, Halifax, NS, Canada Dr. Alex Pappachen James (Research Fellow) Queensland Micro-nanotechnology center, Griffith University, Australia Dr. T. C. Manjunath HKBK College of Engg., Bangalore, India.

Prof. Elboukhari Mohamed Department of Computer Science, University Mohammed First, Oujda, Morocco

TABLE OF CONTENTS

1. Paper 30111332: A Robust Kernel Descriptor for Finger Spelling Recognition based on RGB-D Information (pp. 1-7) Karla Otiniano-Rodrıguez, Guillermo Camara-Chavez Department of Computer Science (DECOM), Federal University of Ouro Preto, Ouro Preto-MG-Brazil Abstract — Systems of communication based on sign language and finger spelling are used by deaf people. Finger spelling is a system where each letter of the alphabet is represented by a unique and discrete movement of the hand. Intensity and depth images can be used to characterize hand shapes corresponding to letters of the alphabet. The advantage of depth sensors over color cameras for sign language recognition is that depth maps provide 3D information of the hand. In this paper, we propose a robust model for finger spelling recognition based on RGB-D information using a kernel descriptor. In the first stage, motivated by the performance of kernel based features, we decided to use the gradient kernel descriptor for feature extraction from depth and intensity images. Then, in the second stage, the Bag-of-Visual-Words approach is used to search semantic information. Finally, the features obtained are used as input of our Support Vector Machine (SVM) classifier. The performance of this approach is quantitatively and qualitatively evaluated on a dataset of real images of the American Sign Language (ASL) finger spelling. This dataset is composed of 120,000 images. Different experiments were performed using a combination of intensity and depth information. Our approach achieved a high recognition rate with a small number of training samples. With 10% of samples, we achieved an accuracy rate of 88.54% and with 50% of samples, we achieved a 96.77%; outperforming other state-of-the-art methods, proving its robustness. 2. Paper 30111304: A Novel Non-Shannon Edge Detection Algorithm for Noisy Images (pp. 8-13) El-Owny, Hassan Badry Mohamed A. Department of Mathematics, Faculty of Science ,Aswan University , 81528 Aswan, Egypt. Current: CIT College, Taif University, 21974 Taif, KSA. Abstract— Edge detection is an important preprocessing step in image analysis. Successful results of image analysis extremely depend on edge detection. Up to now several edge detection methods have been developed such as Prewitt, Sobel, Zerocrossing, Canny, etc. But, they are sensitive to noise. This paper proposes a novel edge detection algorithm for images corrupted with noise. The algorithm finds the edges by eliminating the noise from the image so that the correct edges are determined. The edges of the noise image are determined using non-Shannon measures of entropy. The proposed method is tested under noisy conditions on several images and also compared with conventional edge detectors such as Sobel and Canny edge detector. Experimental results reveal that the proposed method exhibits better performance and may efficiently be used for the detection of edges in images corrupted by Salt-and-Pepper noise. Keywords -Non-Shannon Entropy; Edge Detection; Threshold Value; Noisy images. 3. Paper 30111308: Influence of Stimuli Color and Comparison of SVM and ANN classifier Models for BCI based Applications using SSVEPs (pp. 14-22) Rajesh Singla, Department of Instrumentation and Control Engineering, Dr. B. R. Ambedkar National Institute of Technology Jalandhar, Punjab-144011, India Arun Khosla, Department of Electronics and Communication Engineering, Dr. B. R. Ambedkar National Institute of Technology Jalandhar, Punjab-144011, India Rameshwar Jha, Director General, IET Bhaddal, Distt.- Ropar, Punjab-140108 ,India

Abstract - In recent years, Brain Computer Interface (BCI) systems based on Steady-State Visual Evoked Potential (SSVEP) have received much attentions. In this study four different flickering frequencies in low frequency region were used to elicit the SSVEPs and were displayed on a Liquid Crystal Display (LCD) monitor using LabVIEW. Four stimuli colors, green, blue, red and violet were used in this study to investigate the color influence in SSVEPs. The Electroencephalogram (EEG) signals recorded from the occipital region were segmented into 1 second window and features were extracted by using Fast Fourier Transform (FFT). This study tries to develop a classifier, which can provide higher classification accuracy for multiclass SSVEP data. Support Vector Machines (SVM) is a powerful approach for classification and hence widely used in BCI applications. One-Against-All (OAA), a popular strategy for multiclass SVM is compared with Artificial Neural Network (ANN) models on the basis of SSVEP classifier accuracies. Based on this study, it is found that OAA based SVM classifier can provide a better results than ANN. In color comparison SSVEP with violet color showed higher accuracy than that with other stimuli. Keywords- Steady-State Visual Evoked Potential; Brain Computer Interface; Support Vector Machines; ANN. 4. Paper 30111311: Comparative Study of Person Identification System with Facial Images Using PCA and KPCA Computing Techniques (pp. 23-27) Md. Kamal Uddin, Abul Kalam Azad, Md. Amran Hossen Bhuiyan Department of Computer Science & Telecommunication Engineering, Noakhali Science & Technology University, Noakhali-3814, Bangladesh Abstract — Face recognition is one of the most successful areas of research in computer vision for the application of image analysis and understanding. It has received a considerable attention in recent years both from the industry and the research community. But face recognition is susceptible to variations in pose, light intensity, expression, etc. In this paper, a comparative study of linear (PCA) and nonlinear (KPCA) based approaches for person identification has been explored. The Principal Component Analysis (PCA) is one of the most well-recognized feature extraction tools used in face recognition. The Kernel Principal Component analysis (KPCA) was proposed as a nonlinear extension of a PCA. The basic idea of KPCA is to maps the input space into a feature space via nonlinear mapping and then computes the principal components in that feature space. In this paper, facial images have been classified using Euclidean distance and performance has been analysed for both feature extraction tools. Keywords—Face recognition; Eigenface; Principal component analysis; Kernel principal component analysis. 5. Paper 30111312: Color Image Enhancement of Face Images with Directional Filtering Approach Using Bayer’s Pattern Array (pp. 28-34) Dr. S. Pannirselvam, Research Supervisor & Head, Department of Computer Science, Erode Arts & Science College (Autonomous), Erode, Tamil Nadu, India S. Prasath, Ph.D (Research Scholar), Department of Computer Science, Erode Arts & Science College (Autonomous), Erode, Tamil Nadu, India Abstract - Today, image processing penetrates into various fields, but till it is struggling in quality issues. Hence, image enhancement came into existence as an essential task for all kinds of image processings. Various methods are been presented for color image enhancement, especially for face image. In this paper various filters are used for face image enhancement. In order to improve of the image quality directional filtering approach using Bayer’s pattern are has been applied. In this method the color image are get decomposed into three color component array, then the Bayer’s pattern array is applied to enhance those color component and interpolate the three colors into a single RGB color image. The experimental result shows that this method provides better enhancement in term of quality when compared with the existing methods such as Bilinear Method, Gaussian Filter and Vector Median Filter. The peak Signal Noise Ratio (PSNR) and Mean Square Error (MSE) are been used for similarity measures. Keywords- VMF, GF, BM, PBPM, RGB, YbCr , PSNR, MSE

6. Paper 30111314: An Agent-Based Framework for Virtual Machine Migration in Cloud Computing (pp. 35-39) Somayeh Soltan Baghshahi, Computer Engineering Department, Islamic of Azad University, North Tehran Branch, Tehran, Iran Sam Jabbehdari, Computer Engineering Department, Islamic of Azad University, North Tehran Branch Tehran, Iran Sahar Adabi, Computer Engineering Department, Islamic of Azad University, North Tehran Branch Abstract — Cloud computing is a model for large-scale distributed computing, which services to customers be done through a dynamic virtual resources with high computational power of using the Internet. The cloud service providers use different methods to manage virtual resources, that to use of autonomous nature of the intelligent agents, it can improve quality of service in a cloud distributed environment. In this paper, we design a framework by using of the multiple intelligent agents, which these agent interactions with together and they manage to provide the service. Also, In this framework, an agent is designed to improve the migration technique of virtual machines. Keywords- Cloud Computing; Virtualizaion; Virtual Machine Migration; Agent-Based Framework 7. Paper 30111315: Migration of Virtual Clusters with Using Weighted Fair Queuing Method in Cloud Computing (pp. 40-44) Leila Soltan Baghshahi, Computer Engineering Department, Islamic of Azad University, South Tehran Branch, Tehran, Iran Ahmad Khademzadeh, Education and National International Scientific Cooperation Department, Research Institute for ICT(ITRC), Tehran, Iran Sam Jabbehdari, Computer Engineering Department, Islamic of Azad University, North Tehran Branch, Tehran, Iran Abstract— Load Balancing, Failure Recovery and Quality of Services, portability are some of the advantages in virtualization technology and cloud computing environment. In this environment, with uses the feature of Encapsulation, virtual machines together is considered as a cluster, that these clusters are able to provide the service in cloud environments. In this paper, multiple virtual machines are considered as a cluster. These clusters are migrated from a data center to another data center with using weighted fair queuing. This method is simulated in CloudSim tools in Eclipse and Java programming language. Simulation results show that the bandwidth parameter plays an important role for the virtual machine migration. Keywords-Cloud Computing; Virtualizaion; Virtual Cluster; Live Migration 8. Paper 30111317: Fisher’s Linear Discriminant and Echo State Neural Networks for Identification of Emotions (pp. 45-49) Devi Arumugam, Research Scholar, Department of Computer Science, Mother Teresa Women’s University, Kodaikanal, India. Dr. S. Purushothaman, Professor, PET Engineering College, Vallioor, India-627117. Abstract — Identifying the emotions from facial expression is a fundamental and critical task in human-computer vision. Here expressions like anger, happy, fear, sad, surprise and disgust are identified by Echo State Neural Network. Based on a threshold, the presence of an expression is concluded followed by separation of expression. In each frame, complete face is extracted. The complete face is from top of head to bottom of chin and left ear to right ear. Features are extracted from a face using Fisher’s Linear Discriminant function. The features are extracted from a face is considered as a pattern. If 20 frames belonging to a video are considered, then 20 patterns are created. All 20 patterns are labeled as (1/2/3/4/5/6) according to the labelling decided. The labelling is done as anger=1, fear=2, happy=3, sad=4, surprise=5 and disgust=6. If 20 frames from each video is obtained then number of patterns available for training the proposed Echo State neural Networks are 6 videos x 20 frames= 120 frames. Hence, 120

patterns are formed which are used for training ESNN to obtain final weights. This process is called during the testing of ESNN. In testing of ESNN, FLD features are presented to the input layer of ESNN. The output obtained in the output layer of ANN is compared with threshold to decide the type of expression. For ESNN, the expression identification is highest. Keywords- Video frames; Facial tracking; Eigen Value and eigen vector; Fisher’s Linear Discriminant (FLD); Echo State Neural Network (ESNN); 9. Paper 30111321: A New Current-Mode Multifunction Inverse Filter Using CDBAs (pp. 50-52) Anisur Rehman Nasir, Syed Naseem Ahmad Dept. of Electronics and Communication Engg. Jamia Millia Islamia, New Delhi-110025, India Abstract - A novel current-mode multifunction inverse filter configuration using current differencing buffered amplifiers (CDBAs) is presented. The proposed filter employs two CDBAs and passive components. The proposed circuit realizes inverse lowpass, inverse bandpass and inverse highpass filter functions with proper selection of admittances. The feasibility of the proposed multifunction inverse filter has been tested by simulation program. Simulation results agree well with the theoretical results. Keywords: CDBA, multifunction, inverse filter 10. Paper 30111324: Assessment of Customer Credit through Combined Clustering of Artificial Neural Networks, Genetics Algorithm and Bayesian Probabilities (pp. 53-57) Reza Mortezapour, Department of Electronic And Computer, Islamic Azad University, Zanjan, Iran Mehdi Afzali, Department of Electronic And Computer, Islamic Azad University, Zanjan, Iran Abstract — Today, with respect to the increasing growth of demand to get credit from the customers of banks and finance and credit institutions, using an effective and efficient method to decrease the risk of non-repayment of credit given is very necessary. Assessment of customers' credit is one of the most important and the most essential duties of banks and institutions, and if an error occurs in this field, it would leads to the great losses for banks and institutions. Thus, using the predicting computer systems has been significantly progressed in recent decades. The data that are provided to the credit institutions' managers help them to make a straight decision for giving the credit or not-giving it. In this paper, we will assess the customer credit through a combined classification using artificial neural networks, genetics algorithm and Bayesian probabilities simultaneously, and the results obtained from three methods mentioned above would be used to achieve an appropriate and final result. We use the K_folds cross validation test in order to assess the method and finally, we compare the proposed method with the methods such as Clustering-Launched Classification (CLC), Support Vector Machine (SVM) as well as GA+SVM where the genetics algorithm has been used to improve them. Keywords - Data classification; Combined Clustring; Artificial Neural Networks; Genetics Algorithm; Bayesian Probabilities. 11. Paper 30111327: A Cross Layer UDP-IP protocol for Efficient Congestion Control in Wireless Networks (pp. 58-68) Uma S V, K S Gurumurthy Department of ECE, University Visveswaraya College of Engineering, Bangalore University, Bangalore, India Abstract — Unlike static wired networks, mobile wireless networks present a big challenge to congestion and flow control algorithms as wireless links are in a constant competition to access the shared radio medium. The transport layer along with IP layer plays a major role in Congestion control applications in all such networks. In this research, a twofold approach is used for more efficient Congestion Control. First, a Dual bit Congestion Control Protocol

(DBCC) that uses two ECN bits in the IP header of a pair of packets as feedback is used. This approach differentiates between the error and congestion-caused losses, and is therefore capable of operating in all wireless environments including encrypted wireless networks. Secondly, for better QoS and fairshare of bandwidth in mobile multimedia wireless networks, a combined mechanism, called the Proportional and Derivative algorithm [PDA] is proposed at the transport layer for UDP traffic congestion control. This approach relies on the buffer occupancy to compute the supported rate by a router on the connection path, carries back this information to the traffic source to adapt its actual transmission rate to the network conditions. The PDA algorithm can be implemented at the transport layer of the base station in order to ensure a fair share of the 802.11 bandwidth between the different UDP-based flows. We demonstrate the performance improvements of the cross layer approach as compared to DPCP and VCP through simulation and also the effectiveness of the combined strategy in reducing Network Congestion. Keywords — congestion; explicit congestion bits [ECN]; transport layer; Internet Protocol [IP]; transmission rate; 12. Paper 30111331: The Development of Educational Quality Administration: a Case of Technical College in Southern Thailand (pp. 69-72) Bangsuk Jantawan, Department of Tropical Agriculture and International Cooperation, National Pingtung University of Science and Technology, Pingtung, Taiwan Cheng-Fa Tsai, Department of Management Information Systems, National Pingtung University of Science and Technology, Pingtung, Taiwan Abstract— The purpose of this research were: to survey the needs of using the information system for educational quality administration; to develop Information System for Educational quality Administration (ISEs) in accordance with quality assessment standard; to study the qualification of ISEs; and to study satisfaction level of ISEs user. Subsequently, the tools of study have been employed that there were the collection of 47 questionnaires and 5 interviews to specialist by responsible officers for Information center of Technical colleges and Vocational colleges in Southern Thailand. The analysis of quantitative data has employed descriptive statistics using mean and standard deviation as the tool of measurement. Hence, the result was found that most users required software to search information rapidly (82.89%), software for collecting data (80.85%) and required Information system which could print document rapidly and ready for use (78.72%). The ISEs was created and developed by using Microsoft Access 2007 and Visual Basic. The ISEs was at good level with the average of 4.49 and SD at 0.5. Users’ satisfaction of this software was at good level with the average of 4.36 and SD at 0.58. Keywords- Educational Quality Assurance; Educational Quality Administration; Information System; 13. Paper 31101306: Performance Evaluation Of Data Compression Techniques Versus Different Types Of Data (pp. 73-78) Doa'a Saad El-Shora, Faculty of Computers and Informatics, Zagazig University, Zagazig, Egypt Ehab Rushdy Mohamed, Faculty of Computers and Informatics, Zagazig University, Zagazig, Egypt Nabil Aly Lashin, Faculty of Computers and Informatics, Zagazig University, Zagazig, Egypt Ibrahim Mahmoud El- Henawy, Faculty of Computers and Informatics, Zagazig University, Zagazig, Egypt Abstract — Data Compression plays an important role in the age of information technology. It is now very important a part of everyday life. Data compression has an important application in the areas of file storage and distributed systems. Because real world files usually are quit redundant, compression can often reduce the file sizes considerably, this in turn reduces the needed storage size and transfer channel capacity. This paper surveys a variety of data compression techniques spanning almost fifty years of research. This work illustrates how the performance of data compression techniques is varied when applying on different types of data. In this work the data compression techniques: Huffman, Adaptive Huffman and arithmetic, LZ77, LZW, LZSS, LZHUF, LZARI and PPM are tested against different types of data with different sizes. A framework for evaluation the performance is constructed and applied to these data compression techniques.

(IJCSIS) International Journal of Computer Science and Information Security,Vol. 11, No. 12, December 2013

A Robust Kernel Descriptor for Finger SpellingRecognition Based on RGB-D Information

Karla Otiniano-Rodrıguez #1, Guillermo Camara-Chavez #2

# Department of Computer Science (DECOM), Federal University of Ouro PretoOuro Preto-MG-Brazil

1 [email protected] [email protected]

Abstract—Systems of communication based on sign languageand finger spelling are used by deaf people. Finger spelling isa system where each letter of the alphabet is represented by aunique and discrete movement of the hand. Intensity and depthimages can be used to characterize hand shapes correspondingto letters of the alphabet. The advantage of depth sensors overcolor cameras for sign language recognition is that depth mapsprovide 3D information of the hand. In this paper, we proposea robust model for finger spelling recognition based on RGB-Dinformation using a kernel descriptor. In the first stage, motivatedby the performance of kernel based features, we decided to usethe gradient kernel descriptor for feature extraction from depthand intensity images. Then, in the second stage, the Bag-of-Visual-Words approach is used to search semantic information.Finally, the features obtained are used as input of our SupportVector Machine (SVM) classifier. The performance of this ap-proach is quantitatively and qualitatively evaluated on a datasetof real images of the American Sign Language (ASL) fingerspelling. This dataset is composed of 120,000 images. Differentexperiments were performed using a combination of intensity anddepth information. Our approach achieved a high recognitionrate with a small number of training samples. With 10% ofsamples, we achieved an accuracy rate of 88.54% and with 50%of samples, we achieved a 96.77%; outperforming other state-of-the-art methods, proving its robustness.

I. INTRODUCTION

Sign language is a complex way of communication inwhich hands, limbs, head and facial expression are used tocommunicate a visual-spatial language without sound, mostlyused between deaf people. Deaf people use systems of com-munication based on sign language and finger spelling. Insign language, the basic units are composed by a finite set ofhand configurations, spatial locations and movements. Theircomplex spatial grammars are remarkably different from thegrammars of spoken languages [1], [2]. Hundreds of signlanguages, such as ASL (American Sign Language), BSL(British Sign Language), Auslan (Australian Sign Language)and LIBRAS (Brazilian Sign Language) [1], are in use aroundthe world and are at the cores of local deaf cultures. Unfortu-nately, these languages are barely known outside of the deafcommunity, meaning a communication barrier.

Finger spelling is a system where each letter of the alphabetis represented by a unique and discrete movement of the hand.

Finger spelling integrates a sign language due to many reasons:when a concept lacks a specific sign, for proper nouns, for loansigns (signs borrowed from other languages) or when a sign isambiguous [3]. Each sign language has its own finger spellingsimilar to different characters in different languages.

Several techniques have been developed to achieve anadequate recognition rate of sign language. Over the yearsand with the advance of technology, methods have been pro-posed in order to improve the data acquisition, processing orclassification, such is the case of image acquisition. There arethree main approaches: sensor-based, vision-based and hybridsystems using a combination of these systems. Sensor-basedmethods use sensory gloves and motion trackers to detect handshapes and body movements. Vision-based methods, that usestandard cameras, image processing, and feature extraction,are used for capturing and classifying hand shapes and bodymovements. Hybrid systems use information from vision-based camera and other type of sensors like infrared depthsensors.

Sensor-based methods, such as data gloves, can provide ac-curate measurements of hands and movement. Unfortunately,these methods require extensive calibration, they also restrictthe natural movement of hands and are often very expensive.Video-based methods are less intrusive, but new problemsarise: locating the hands and segmenting them is a non-trivial task. Recently, depth cameras have become popularat a commodity price. Depth information makes the task ofsegmenting the hand from the background much easier. Depthinformation can be used to improve the segmentation process,as used in [4], [5], [6], [7].

Recently, depth cameras have raised a great interest in thecomputer vision community due to their success in manyapplications, such as pose estimation [8], [9], tracking [10],object recognition [10], etc. Depth cameras were also used forhand gesture recognition [11], [12], [13], [14], [15]. Uebersaxet al. [12] present a system for recognizing letter and fingerspelled words. Pugeault & Bowden [11] use a MicrosoftKinectTM device to collect RGB and depth images. Theyextracted features using Gabor filters and then a RandomForest predicts the letters from the American Sign Language

1 http://sites.google.com/site/ijcsis/ ISSN 1947-5500


(ASL) finger spelling alphabet. Issacs & Foo [16] proposedan ASL finger spelling recognition system based on neuralnetworks applied to wavelets features. Bergh & Van Gool [17]propose a method based on a concatenation of depth and color-segmented images, using a combination of Haar wavelets andneural networks for 6 hand poses recognition of a single user.

In this paper, we propose a framework for finger spellingrecognition using intensity and depth images. Motivated bythe performance of kernel based features, due to its simplicityand the ability to turn any type of pixel attribute into patch-level features, we decided to use the gradient kernel descriptor[18]. The experiments are performed using a public databasecomposed of 120,000 images stating 24 symbols classes [19].The obtained results show that the accuracy obtained by ourmethod, using intensity and depth images, is greater thanonly using intensity or depth images separately. Moreover, theaccuracy obtained by the proposed method performs betterthan the methods proposed in [11], [15]. The results showthat our method is promising.

The remainder of this paper is organized as follows. InSection II, our proposed method is introduced and detailed.The experiments are presented in Section III, where theresults are discussed. Finally, conclusion and future work arepresented in Section IV.

II. PROPOSED MODEL

This section describes the methodology developed to per-form a finger spelling recognition from RGB-D information.The proposed model consists of two stages as shown inFigure 1. In the first stage, we apply the bag-of-visual-words approach, this technique consists of three steps, featuredescription, vocabulary generation and histogram generation.For feature extraction, we use intensity and depth images andthe gradient kernel descriptor is applied on those images. Thiskernel descriptor consists of three kernels. The normalized lin-ear kernel weighs the contribution of each pixel using gradientmagnitudes, an orientation kernel computes the similarity ofgradient orientations and finally a position Gaussian kernelmeasures how close two pixels are spatially. The grouping bysimilarity of features extracted in the previous step generatesthe visual vocabulary, the centroid of each group representsa visual word. Thus, the visual words histogram is obtainedby counting the number of occurrences of each visual word.Finally, in the second stage, these histograms are used as inputto our SVM classifier.

A. Bag-of-Visual-Words

Bag-of-Visual-Words has first been introduced by Sivic forvideo retrieval [20]. Due to its efficiency and effectiveness,it became very popular in the fields of image retrieval andcategorization. Image categorization techniques rely either onunsupervised or supervised learning.

Our model uses the Bag-of-Visual-Words approach in orderto search semantic information. The original method workswith documents and words. Therefore, we consider an image

Fig. 1. Proposed model for finger spelling recognition.

as a document and the ”words” will be the visual entities foundin the image. The Bag-of-Visual-Words approach consists ofthree operations: feature description, visual word vocabularygeneration and histogram generation.

1) Feature Description: Gradient Kernel Descriptor: Thelow-level image feature extractor, kernel descriptor, designedfor visual recognition in [21], consists of three steps: designmatch kernel using some pixel attribute, learn compact ba-sis vectors using Kernel Principle Component Analysis andconstruct kernel (KPCA) descriptor by projecting the infinite-dimensional feature vector to the learned basis vectors. Theauthors proposed three types of effective kernel descriptorsusing gradient, color and shape pixel attributes. In othermodel proposed by the same authors [18], the gradient kerneldescriptor is applied over depth images. Thereby, in order tocapture edge cues in depth maps, we used the gradient matchkernel, Kgrad :

Kgrad(P,Q) =∑p∈P

∑q∈Q

m(p)m(q)ko(θ(p), θ(q))ks(p, q)

(1)The normalized linear kernel m(p)m(q) weighs

the contribution of each gradient where m(p) =

m(p)/√∑

p∈P m(p)2 + εg and εg is a small positiveconstant to ensure that the denominator is larger than 0and m(p) is the magnitude of the depth gradient at a pixelp. Then, ko(θ(p), θ(q)) = exp(−γo‖θ(p) − θ(q)‖2) is aGaussian kernel over orientations. The authors [21] suggestto set γo = 5. To estimate the difference between orientationsat pixels p and q, we use the following normalized gradient



vectors in the kernel function ko:

θ(p) = [sin(θ(p))cos(θ(p))]

θ(q) = [sin(θ(q))cos(θ(q))]

where θ(p) is the orientation of the depth gradient at a pixelp. Gaussian position kernel ks(p, q) = exp(−γs‖p − q‖2)with p denoting the 2D position of a pixel in an imagepatch (normalized to [0,1]), measures how close two pixelsare spatially. The value suggest for γs is 3.

To summarize, the gradient match kernel Kgrad consists ofthree kernels: the normalized linear kernel weighs the contri-bution of each pixel using gradient magnitudes; the orientationkernel ko computes the similarity of gradient orientations; andthe position Gaussian kernel ks measures how close two pixelsare spatially.

Match kernels provide a principled way to measure thesimilarity of image patches, but evaluating kernels can becomputationally expensive when image patch are large [21].The corresponding kernel descriptor can be extracted from thismatch kernel by projecting the infinite-dimensional featurevector to a set of finite basis vectors, which are the edgefeatures that we use in the next steps. For more details, theapproach that extracts the compact low-dimensional featuresfrom match kernels is found in [21].

2) Vocabulary Generation: Then, a visual word vocabularyis generated from the feature vectors;s each visual word(codeword) represents a group of several similar features.The visual word vocabulary (codebook) defines a space ofall entities occurring in the image.

3) Histogram Generation: Finally, a histogram of visualwords is created by counting the occurrence of each codeword.These occurrences are counted and arranged in a vector. Eachvector represents the features for an image.

B. Classification

Support vector machines, introduced as a machine learningmethod by Cortes and Vapnik [22], are a useful classificationmethod. Furthermore, SVMs have been successfully appliedin many real world problems and in several areas: text cate-gorization, handwritten digit recognition, object recognition,etc. The SVMs have been developed as a robust tool forclassification and regression in noisy and complex domains.SVM can be used to extract valuable information from datasets and construct fast classification algorithms for massivedata.

An important characteristic of the SVM classifier is to allowa non-linear classification without requiring explicitly a non-linear algorithm thanks to kernel theory.

In kernel framework data points may be mapped into ahigher dimensional feature space, where a separating hyper-plane can be found. We can avoid to explicitly computingthe mapping using the kernel trick which evaluate similar-ities between data K(dt, ds) in the input space. Commonkernel functions are: linear, polynomial, Radial Basis Function(RBF), χ2 distance and triangular.

Fig. 3. Most conflictive similar signs in the dataset.

III. EXPERIMENTS

The ASL Finger Spelling Dataset [19] contains 500 samplesfor each of 24 signs, recorded from 5 different persons (non-native to sign language), amounting to a total of 60,000samples. Each sample has a RGB image and a depth image,making a total of 120,000 images. The sign J and Z arenot used, because these signs have motion and the proposedmodel only works with static signs. The dataset has varietyof background and viewing angles. Figure 2 shows someexamples and there is possible to see the variety in size,background and orientation.

Due to the variety in the orientation when the signal isperformed, signs became strongly similar. Figure 3 shows themost similar signs a, e, m, n, s and t. The examples are takenfrom the same user. It is easy to identify the similarity betweenthese signs, all are represented by a closed fist, and differonly by the thumb position, leading to higher confusion levels.Therefore, these signs are the most difficult to differentiate inthe classification task.

In order to validate our technique, we conduct three experi-ments. In the first, a classification of the signs was performedusing different percentages of samples for training and testingfrom intensity information. In the second, a classification wasalso performed from depth information varying the percent-ages of training and testing. Finally, a classification of thesigns was performed using different percentages of samplesfor training and testing from both information (RGB-D).

For each experiment, we have some specifications:• To extract all low level features using gradient kernel

descriptor, are used approximately 12x13 patches overdense regular grid with spacing of 8 pixel (images arenot of uniform size), each patch has a size of 16x16.

• In order to produce the visual word vocabulary, the LBG(Linde-Buzo-Gray) [23] algorithm was used to detect onehundred clusters by taking a sample of 30% from the totalfeatures.

• Moreover, in the classification stage, we use a RBFkernel, whose values for g (gamma) and c (cost) are 0.25and 5, respectively. We also use different percentages ofsamples for training and testing. For example, we use10% of samples for training and the other 90% is used totesting, and this percentage varies up to 50% for training.In order to obtain more precise results, each experiment



Fig. 2. ASL Finger Spelling Dataset: 24 static signs by 5 users. It is an example of the variety of the dataset. This array shows one image from each userand from each letter.

TABLE IACCURACIES AND STANDARD DEVIATION OF THE CLASSIFICATION USING

INTENSITY INFORMATION.

% Training % Testing Accuracy Standard deviation10 90 79.08% 0.2520 80 85.28% 0.2130 70 88.32% 0.2040 60 90.19% 0.1550 50 91.58% 0.16

was performed 30 times and we show the mean accuracyfor each one. The library LIBSVM (a library for SupportVector Machines)] [24] was used in our implementation.First experiment: An average accuracy of 79.08% when a

10% of samples are used for training and 90% for testing. Thisaccuracy is the mean of the values of the main diagonal of theconfusion matrix and represents the signs correctly classified(true positives). This accuracy increases when more samplesare used for training. With 30% of samples for training isobtained 88.32% and when 50% of samples are training weobtain 91.58% of accuracy. More results are found in the TableI. The classification using intensity information was improvedcompared to the proposed model found in [15], in which,was obtained an accuracy of 62.70% using the same type ofinformation.

Second experiment: For this experiment, using depthinformation, the average accuracy obtained was 75.6% when10% of samples are used for training. The higher accuracy,86.86%, was obtained using 50% of samples for training andthe other 50% of samples for testing. Other results are foundin the Table II. This results show a slight increase in theclassification rate compared to the results found in [15], wherewas obtained an accuracy of 85.18%.

Third experiment: The classification task was performedusing RGB-D information. The data for this experiment wasobtained by joining the features (histograms) from RGB anddepth information, which were used in the experiments 1 and

TABLE IIACCURACIES AND STANDARD DEVIATION OF THE CLASSIFICATION USING

DEPTH INFORMATION


2, respectively. Is obtained an average accuracy of 96.77%when 50% of samples are used for training. In other case, whenare used 10% of samples for training, we obtain an accuracyof 88.54%. It means that are used 250 samples for each signfor training and 2250 samples for testing. Table III shows theresults for this case (10% to training). Signs f, b, l and yhave the highest average accuracies (over 95%). Otherwise, thesigns n, m, r and t have the lowest values (with 80% and 81%).The low recognition value of sign t is due to the big similaritywith signs m and n, as shown in the Figure 3. Table IV showsthe results when 50% of samples are used for training. In thesimilar case, the signs with highest accuracies a, b, f and lhave 99% of recognition. Otherwise, signs t, v, m and n havean accuracy between 93% and 94%. However, each sign havean accuracy greater than 93%, proving the high recognitionrate of our proposed model. In Table V are found the averageaccuracies for each experiment using different percentages ofsamples for training.

We summarize and compare the results in Tables VI andVII. It includes the average accuracy and standard deviation foreach experiment. We can see that using RGB-D informationwe obtain the highest average accuracy, outperforming theintensity and depth methods and also the methods proposed byPugeault & Bowden [11] and Zhu & Wong [14]. These lastmethods are found in the state-of-the-art and use the samedataset, the principal difference between these methods is thenumber of samples used for training. Pugeault & Bowden [11]



TABLE IIICONFUSION MATRIX OF THE CLASSIFICATION OF 24 SIGN USING RGB-D INFORMATION WITH 10% FOR TRAINING AND 90% FOR TESTING.

a b c d e f g h i k l m n o p q r s t u v w x ya 0.94 0.00 0.00 0.00 0.00 0.00 0.01 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.02 0.00 0.00 0.00 0.00 0.00b 0.01 0.95 0.01 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.01 0.00 0.00 0.00 0.00c 0.01 0.01 0.93 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00d 0.00 0.00 0.00 0.88 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.01 0.00 0.00 0.02 0.00 0.00 0.01 0.01 0.00 0.01 0.00e 0.03 0.00 0.01 0.01 0.85 0.00 0.00 0.00 0.01 0.01 0.01 0.01 0.01 0.02 0.00 0.00 0.00 0.01 0.01 0.00 0.00 0.00 0.01 0.00f 0.00 0.00 0.00 0.00 0.00 0.97 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00g 0.01 0.00 0.00 0.00 0.00 0.00 0.91 0.05 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00h 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.94 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00i 0.01 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.92 0.01 0.01 0.01 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.01k 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.01 0.88 0.00 0.00 0.01 0.00 0.00 0.00 0.02 0.00 0.01 0.02 0.02 0.00 0.01 0.00l 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.95 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01

m 0.01 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.01 0.00 0.81 0.09 0.00 0.00 0.00 0.01 0.03 0.01 0.00 0.00 0.00 0.00 0.00n 0.01 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.01 0.00 0.06 0.81 0.01 0.00 0.00 0.00 0.02 0.03 0.00 0.00 0.00 0.02 0.00o 0.00 0.00 0.01 0.02 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.01 0.85 0.01 0.01 0.00 0.02 0.01 0.00 0.00 0.00 0.01 0.00p 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.92 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00q 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.92 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00e 0.01 0.01 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.01 0.00 0.00 0.00 0.80 0.00 0.00 0.04 0.06 0.00 0.02 0.00s 0.01 0.00 0.00 0.00 0.03 0.00 0.00 0.00 0.01 0.00 0.00 0.03 0.05 0.01 0.00 0.00 0.00 0.83 0.01 0.00 0.00 0.00 0.01 0.00t 0.02 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.01 0.00 0.03 0.08 0.01 0.00 0.00 0.00 0.02 0.80 0.00 0.00 0.00 0.01 0.00u 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.01 0.00 0.00 0.00 0.05 0.00 0.00 0.85 0.04 0.02 0.00 0.00v 0.01 0.01 0.00 0.01 0.01 0.01 0.00 0.00 0.00 0.01 0.00 0.00 0.01 0.00 0.00 0.00 0.04 0.00 0.00 0.05 0.82 0.03 0.00 0.00w 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.01 0.03 0.93 0.00 0.00x 0.00 0.00 0.01 0.02 0.01 0.01 0.00 0.00 0.00 0.01 0.00 0.00 0.02 0.00 0.01 0.01 0.01 0.00 0.01 0.00 0.00 0.00 0.86 0.00y 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.95

TABLE IVCONFUSION MATRIX OF THE CLASSIFICATION OF 24 SIGN USING RGB-D INFORMATION WITH 50% OF SAMPLES FOR TRAINING AND 50% FOR TESTING.

a b c d e f g h i k l m n o p q r s t u v w x ya 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00b 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00c 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00d 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00e 0.00 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00f 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00g 0.00 0.00 0.00 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00h 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00i 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00k 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00l 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

m 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.94 0.03 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00n 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.94 0.01 0.00 0.00 0.00 0.01 0.01 0.00 0.00 0.00 0.00 0.00o 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.97 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00p 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.98 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00q 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00r 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.94 0.00 0.00 0.02 0.01 0.00 0.00 0.00s 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.01 0.01 0.00 0.00 0.00 0.96 0.01 0.00 0.00 0.00 0.00 0.00t 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.02 0.00 0.00 0.00 0.00 0.01 0.93 0.00 0.00 0.00 0.00 0.00u 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.03 0.00 0.00 0.94 0.01 0.00 0.00 0.00v 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.03 0.93 0.01 0.00 0.00w 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.97 0.00 0.00x 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.97 0.00y 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.98



TABLE VACCURACIES AND STANDARD DEVIATION OF THE CLASSIFICATION USING

RGB-D INFORMATION


TABLE VIACCURACIES AND STANDARD DEVIATION OF THE THREE EXPERIMENTS

USING 10% OF SAMPLES FOR TRAINING.

Method Accuracy Standard DeviationRGB 79.08% 0.25Depth 75.6% 0.26

RGB-D 88.54% 0.17Zhu & Wong [14] 88.9% 0.39

use the 50% of samples (1250 samples) for training and Zhu& Wong [14] use only 40 samples for training. In the TableVI are found results when 10% of samples are used to trainingwith each type of information and the result for the methodproposed by Zhu & Wong [14]. In the Table VII are foundthe accuracies for each experiment when 50% of samples areused for training, also this is the case of the method proposedby Pugeault & Bowden [11].

IV. CONCLUSION AND FUTURE WORK

In this paper, we propose a method for Finger SpellingRecognition from RGB-D information using a robust kerneldescriptor. Then, Bag-of-Visual-Words was applied in orderto search semantic information. Finally, the classification taskis performed by a SVM. The combination of RGB anddepth descriptors obtains the best results (96.77%) with alow variance. Our method achieves a better differentiation ofsimilar signs like n, r and t, incrementing the recognitionrate. The Gradient kernel descriptor has the advantage thatcan be directly applied on the depth images without havingto compute the cloud of points, consequently, reducing thecomputation time. In a previously proposed model [15], weused segmentation to better detect the hand. Even though inthis paper we do not segment the images, we obtain betterresults, showing the robustness of kernel descriptors. As future

TABLE VIIACCURACIES AND STANDARD DEVIATION OF THE THREE EXPERIMENTS

USING 50% OF SAMPLES FOR TRAINING.

Method Accuracy Standard DeviationRGB 91.58% 0.16Depth 86.85% 0.17

RGB-D 96.77% 0.09Pugeault & Bowden[11] 75.00% -

work, we pretend to test other kernels over depth and intensityimages. We also intend to extend our method to recognizedynamic signs.

V. ACKNOWLEDGMENTThe authors are thankful to CNPq, CAPES and FAPEMIG

(Projeto Universal 02292-12), Brazilian funding agencies andto the Federal University of Ouro Preto (UFOP) for supportingthis work.

REFERENCES

[1] LIBRAS, “Brazilian sign language,” http://www.libras.org.br/, last visit:March 10, 2012.

[2] P. W. Vamplew, “Recognition of sign language gestures using neuralnetworks,” Australian Journal of Intelligent Information ProcessingSystems, vol. 5, pp. 27–33, 1996.

[3] A. Puente, J. M. Alvarado, and V. Herrera, “Fingerspelling and signlanguage as alternative codes for reading and writing words for Chileandeaf signers,” American Annals of the Deaf, vol. 151, no. 3, pp. 299–310,2006.

[4] Z. Ren, J. Yuan, and Z. Zhang, “Robust hand gesture recognition basedon finger-earth mover’s distance with a commodity depth camera,” inProceedings of the 19th ACM International Conference on Multimedia.ACM, 2011, pp. 1093–1096.

[5] V. Frati and D. Prattichizzo, “Using Kinect for hand tracking andrendering in wearable haptics,” in Proceedings of the IEEE WorldHaptics Conference (WHC). IEEE, 2011, pp. 317–321.

[6] Y. Li, “Hand gesture recognition using Kinect,” in Proceedings of the3rd IEEE International Conference on Software Engineering and ServiceScience (ICSESS). IEEE, 2012, pp. 196–199.

[7] Z. Mo and U. Neumann, “Real-time hand pose recognition using low-resolution depth images,” in Proceedings of the IEEE Computer SocietyConference on Computer Vision and Pattern Recognition, vol. 2. IEEE,2006, pp. 1499–1505.

[8] G. Fanelli, J. Gall, and L. V. Gool, “Real time head pose estimation withrandom regression forests,” in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR), June 2011, pp. 617–624.

[9] J. Shotton, T. Sharp, A. Kipman, A. W. Fitzgibbon, M. Finocchio,A. Blake, M. Cook, and R. Moore, “Real-time human pose recognitionin parts from single depth images,” Communications of the ACM, vol. 56,no. 1, pp. 116–124, 2013.

[10] I. Oikonomidis, N. Kyriazis, and A. Argyros, “Efficient model-based3D tracking of hand articulations using Kinect,” in Proceedings of theBritish Machine Vision Conference. BMVA Press, 2011, pp. 101.1–101.11.

[11] N. Pugeault and R. Bowden, “Spelling it out: Real-time ASL fin-gerspelling recognition.” in Proceedings of the IEEE InternationalConference on Computer Vision Workshops (ICCV Workshops). IEEE,2011, pp. 1114–1119.

[12] D. Uebersax, J. Gall, M. V. den Bergh, and L. J. V. Gool, “Real-time signlanguage letter and word recognition from depth data,” in Proceedingsof the IEEE International Conference on Computer Vision Workshops(ICCV Workshops), 2011, pp. 383–390.

[13] M. d. S. Anjo, E. B. Pizzolato, and S. Feuerstack, “A real-time systemto recognize static gestures of brazilian sign language (libras) alphabetusing kinect,” in Proceedings of the 11th Brazilian Symposium onHuman Factors in Computing Systems. Brazilian Computer Society,2012, pp. 259–268.

[14] X. Zhu and K.-Y. K. Wong, “Single-frame hand gesture recognitionusing color and depth kernel descriptors,” in Proceedings of the 21stInternational Conference on Pattern Recognition (ICPR). IEEE, 2012,pp. 2989–2992.

[15] K. Otiniano-Rodrıguez and G. Camara-Chavez, “Finger spelling recog-nition from RGB-D information using kernel descriptor,” in Proceedingsof the SIBGRAPI 2013 (XXVI Conference on Graphics, Patterns andImages), 2013.

[16] J. Isaacs and S. Foo, “Hand pose estimation for american sign languagerecognition,” 36th Southeastern Symposium on System Theory, pp. 132–136, 2004.



[17] M. Van den Bergh and L. Van Gool, “Combining RGB and ToF camerasfor real-time 3D hand gesture interaction,” in Proceedings of the IEEEWorkshop on Applications of Computer Vision (WACV), ser. WACV ’11.Washington, DC, USA: IEEE Computer Society, 2011, pp. 66–72.

[18] L. Bo, X. Ren, and D. Fox, “Depth kernel descriptors for objectrecognition,” in Proceedings of the IEEE International Conference onIntelligent Robots and Systems (IROS). IEEE, 2011, pp. 821–826.

[19] R. B. Nicolas Pugeault, “ASL finger spelling dataset,”http://personal.ee.surrey.ac.uk/Personal/N.Pugeault/index.php, lastvisit: April 29, 2013.

[20] J. Sivic and A. Zisserman, “Video google: A text retrieval approachto object matching in videos,” in Proceedings of the Ninth IEEEInternational Conference on Computer Vision. IEEE, 2003, pp. 1470–1477.

[21] L. Bo, X. Ren, and D. Fox, “Kernel descriptors for visual recognition,”Advances in Neural Information Processing Systems, vol. 7, 2010.

[22] C. Cortes and V. Vapnik, “Support-vector networks,” Machine Learning,vol. 20, no. 3, pp. 273–297, 1995.

[23] Y. Linde, A. Buzo, and R. Gray, “An algorithm for vector quantizerdesign,” IEEE Transactions on Communications, vol. 28, no. 1, pp. 84–95, 1980.

[24] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for supportvector machines,” ACM Transactions on Intelligent Systems andTechnology, vol. 2, no. 3, pp. 1–27, 2011, software available athttp://www.csie.ntu.edu.tw/ cjlin/libsvm.


(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, 2013

A novel non-Shannon edge detection algorithm for noisy images

El-Owny, Hassan Badry Mohamed A.

Department of Mathematics, Faculty of Science ,Aswan University , 81528 Aswan, Egypt. Current: CIT College, Taif University, 21974 Taif, KSA.

.

Abstract— Edge detection is an important preprocessing step in image analysis. Successful results of image analysis extremely depend on edge detection. Up to now several edge detection methods have been developed such as Prewitt, Sobel, Zero-crossing, Canny, etc. But, they are sensitive to noise. This paper proposes a novel edge detection algorithm for images corrupted with noise. The algorithm finds the edges by eliminating the noise from the image so that the correct edges are determined. The edges of the noise image are determined using non-Shannon measures of entropy. The proposed method is tested under noisy conditions on several images and also compared with conventional edge detectors such as Sobel and Canny edge detector. Experimental results reveal that the proposed method exhibits better performance and may efficiently be used for the detection of edges in images corrupted by Salt-and-Pepper noise. Keywords -Non-Shannon Entropy ; Edge Detection; Threshold Value ; Noisy images.

I. INTRODUCTION

Edge detection has been used extensively in areas related to image and signal processing. Its use includes pattern recognition, image segmentation, and scene analysis. The edges are also use to locate the objects in an image and measure their geometrical features. Hence, edge detection is an important identification and classification tool in computer vision. This topic has attracted many researchers and several achievements have been made to investigate new and more robust techniques .

Natural images are prone to noise and artifacts. Salt and pepper noise is a form of noise typically seen on images. It is typically manifested as randomly occurring white and black pixels. Salt and pepper noise creeps into images in situations where quick transients, such as faulty switching, take place. On the other hand, White noise is additive in nature where the each pixel in the image is modified via the addition of a value drawn from a Gaussian distribution. To test the generality of the results, the proposed edge detection algorithm was tested on images containing both these types of noise.

A large number of studies have been published in the field of image edge detection[1-16], which attests to its importance within the field of image processing. Many edge detection algorithms have been proposed, each of which has its own strengths and weaknesses; for this reason, hitherto there does not appear to be a single "best" edge detector. A good edge

detector should be able to detect the edge for any type of image and should show higher resistance to noise.

Examples of approaches to edge detection include algorithms such as the Sobel and Prewitt edge detectors which are based on the first order derivative of the pixel intensities[1]. The Laplacian-of-Gaussian (LoG) edge detector is another popular technique, using instead the second order differential operators to detect the location of edges [2,17,18]. However, all of these algorithms tend to be sensitive to noise, which is an intrinsically high frequency phenomenon. To solve this problem the Canny edge detector was proposed, which combines a smoothing function with zero crossing based edge detection [3]. Although it is more resilient to noise than the previously mentioned algorithms, its performance is still not satisfactory when the noise level is high. There are many situations where sharp changes in color intensity do not correspond to object boundaries like surface marking, recording noise and uneven lighting conditions [4-7,19-22]

In this paper we present a new approach to detect edges of gray scale noisy images based on information theory, which is entropy based thresholding. The proposed method is decrease the computation time as possible as can and the results were very good compared with the other methods.

The paper is organized as follows: Section 2 describes in brief the basic concepts of Shannon and non-Shannon entropies. Section 3 is devoted to the proposed method of edge detection. In Section 4, the details of the edge detection algorithm is described. In Section 5, some particular images will be analyzed using proposed method based algorithm and moreover, a comparison with some existing methods will be provided for these images. Finally, conclusions will be drawn in Section 6.

II. BASIC CONCEPT OF ENTROPY

Physically Entropy can be associated with the amount of disorder in a physical system. In[23] Shannon redefined the entropy concept of Boltzmann/Gibbs as a measure of uncertainty regarding the information content of a system. He defined an expression for measuring quantitatively the amount of information produced by a process.

In accordance with this definition, a random event that occurs with probability is said to contain ln 1⁄ ln units of information. The amount

is called the self-information of event . The amount of



self information of the event is inversely proportional to its probability. If 1, then 0 and no information is attributed to it. In this case, uncertainty associated with the event is zero. Thus, if the event always occurs, then no information would be transferred by communicating that the event has occurred. If 0.8 , then some information would be transferred by communicating that the event has occurred[24].

The basic concept of entropy in information theory has to do with how much randomness is in a signal or in a random event. An alternative way to look at this is to talk about how much information is carried by the signal [25]. Entropy is a measure of randomness.

Let , , , be the probability distribution of a discrete source. Therefore, 0 1, 1,2, , and ∑ 1, where k is the total number of states. The entropy of a discrete source is often obtained from the probability distribution. The Shannon Entropy can be defined as

ln

This formalism has been shown to be restricted to the domain of validity of the Boltzmann–Gibbs–Shannon (BGS) statistics. These statistics seem to describe nature when the effective microscopic interactions and the microscopic memory are short ranged. Generally, systems that obey BGS statistics are called extensive systems. If we consider that a physical system can be decomposed into two statistical independent subsystems and , the probability of the composite system is , it has been verified that the Shannon entropy has the extensive property (additive):

(1) From [25] , (2) where ψ α is a function of the entropic index. In Shannon entropy 1.

Rènyi entropy[26] for the generalized distribution can be written as follows:

11

ln , 0 ,

this expression meets the BGS entropy in the limit 1. Rènyi entropy has a nonextensive property for statistical independent systems, defined by the following pseudo additivity entropic formula

1 . Tsallis[27-29] has proposed a generalization of the BGS statistics, and it is based on a generalized entropic form,

∑ , where k is the total number of possibilities of the system and the real number α is an entropic index that characterizes the degree of nonextensivity. This expression meets the BGS entropy in the limit 1 . The Tsallis entropy is nonextensive in such a way that for a statistical independent

system, the entropy of the system is defined by the following pseudo additive entropic rule

1

The generalized entropies of Kapur of order α and type β [30,31] is

, ln ∑

∑, , , 0 (3)

In the limiting case, when α 1 and β 1, H , p reduces to and when β 1, H , p reduces to . Also, H , p is a composite function which satisfies pseudo-additivity as:

, , , 1 ,

, . (4)

III. SELECTION OF THRESHOLD VALUE BASED ON KAPUR

ENTROPY

A gray level image can be represented by an intensity function, which determines the gray level value for each pixel in the image. Specifically, in a digital image of size an intensity function , , | 1,2, , , 1,2, , , takes as input a particular pixel from the image,

and outputs its gray level value, which is usually in the range of 0 to 255 (if 256 levels are used).

Thresholding produces a new image based on the original one represented by f. It is basically another function , , which produces a new image (i.e. the thresholded image). A threshold is calculated for each pixel value. This threshold is compared with the original image (i.e. ) to determine the new value of the current pixel. can be represented by the following equation [31,32].

,0, if ,1, if , , is the thresholding value.

When Entropy applied to image processing techniques, entropy measures the normality (i.e. normal or abnormal) of a particular gray level distribution of an image. When a whole image is considered, the Kapur entropy as defined in (3) will indicate to what extent the intensity distribution is normal. When we extend this concept to image segmentation, i.e. dealing with foreground(Object) and background regions in an image, the entropy is calculated for both regions, and the subsequent entropy value provides an indication of the normality of the segmentation. In this case, two equations are need for each region, each of them called priori.

In image thresholding, when applying maximum entropy, every gray level value is a candidate to be the threshold value. Each value will be used to classify the pixels into two groups based on their gray levels and their affinity, as less or greater than the threshold value ( ).

Let , , … . , , , … . , be the probability distribution for an image with k gray-levels, where is the normalized histogram i.e. ⁄ and is the gray level histogram. From this distribution, we can derive two



probability distributions, one for the object (class A) and the other for the background (class B), are shown as follows:

: , , … . . , ,

(5)

: , , … . . , ,

where ∑ , ∑ , t is the threshold value. (6)

In terms of the definition of Kapur entropy of order and type , the entropy of Object pixels and the entropy

of background pixels can be defined as follows:

,1

ln∑

∑, , , 0

(7)

,1

ln∑

∑, , , 0 .

The Kapur entropy , is parametrically dependent upon the threshold value for the object and background. It is formulated as the sum each entropy, allowing the pseudo-additive property for statistically independent systems, as defined in (4). We try to maximize the information measure between the two classes (object and background). When

, is maximized, the luminance level that maximizes the function is considered to be the optimum threshold value. This can be achieved with a cheap computational effort.

Argmax , , 1 · , · , . (8) When 1 and 1 , the threshold value in (4),

equals to the same value found by Shannon Entropy. Thus this proposed method includes Shannon’s method as a special case. The following expression can be used as a criterion function to obtain the optimal threshold at 1 and 1.

Argmax , , . (9)

Now, we can describe the Kapur Threshold algorithm to determine a suitable threshold value and α and β as follows:

II. if 0 1 and 0 1 then

Apply Equation (8) to calculate optimum threshold value .

else Apply Equation (9) to calculate optimum threshold value .

end-if 5. Output: The suitable threshold value of I, FOR

, 0,

IV. PROPOSED ALGORITHM FOR EDGE DETECTION

The process of spatial filtering consists simply of moving a filter mask w of order from point to point in an image. At each point , , the response of the filter at that point is calculated a predefined relationship. We will use the usual masks for detection the edges. Assume that 2 1 and

2 1 , where , are nonnegative integers. For this purpose, smallest meaningful size of the mask is 3 3 , as shown in Fig. 1[1].

1, 1 1,0 1, 10, 1 0,0 0,11, 1 1,0 1,1

Fig. 1: Mask coefficients showing coordinate arrangement

1, 1 1, 1, 1 , 1 , , 11, 1 1, 1, 1

Fig. 2

Image region under the above mask is shown in Fig. 2. In order to edge detection, firstly classification of all pixels that satisfy the criterion of homogeneousness, and detection of all pixels on the borders between different homogeneous areas. In the proposed scheme, first create a binary image by choosing a suitable threshold value using Kapur entropy. Window is applied on the binary image. Set all window coefficients equal to 1 except centre, centre equal to × as shown in Fig. 3.

1 1 1 1 × 1 1 1 1

Fig. 3 Move the window on the whole binary image and find the

probability of each central pixel of image under the window. Then, the entropy of each Central Pixel of image under the window is calculated as ln .

TABLE I . P AND H OF CENTRAL PIXEL UNDER WINDOW. 1/9 2/9 3/9 4/9 5/9 6/9 7/9 8/9

0.2441 0.3342 0.3662 0.3604 0.3265 0.2703 0.1955 0.1047

ALGORITHM 1: THRESHOLD VALUE SELECTION (KAPUR

THRESHOLD)

1. Input: A digital grayscale noisy image I of size . 2. Let , be the original gray value of the pixel at the

point , , 1,2, , , 1,2, , 3. Calculate the probability distribution , 0 255 4. For all 0,1, … ,255 ,

I. Apply Equations (5) and (6) to calculate , , and



where, is the probability of central pixel of binary image under the window. When the probability of central pixel

1 then the entropy of this pixel is zero. Thus, if the gray level of all pixels under the window homogeneous, then

1 and 0. In this case, the central pixel is not an edge pixel. Other possibilities of entropy of central pixel under window are shown in Table 1.

In cases 8/9, and 7/9, the diversity for gray level of pixels under the window is low. So, in these cases, central pixel is not an edge pixel. In remaining cases, 6/9, the diversity for gray level of pixels under the window is high. So, for these cases, central pixel is an edge pixel. Thus, the central pixel with entropy greater than and equal to 0.2441 is an edge pixel, otherwise not.

The following Algorithm summarize the proposed technique for calculating the optimal threshold values and the edge detector.

The steps of our proposed technique are as follows: Step 1: Find global threshold value ( ) using Kapur entropy .

The image is segmented by into two parts, the object (Part1) and the background (Part2).

Step 2: By using Kapur entropy, we can select the locals threshold values ( ) and ( ) for Part1 and Part2, respectively.

Step 3: Applying Edge Detection Procedure with threshold values , and . Step 4: Merge the resultant images of step 3 in final output

edge image. In order to reduce the run time of the proposed algorithm,

we make the following steps: Firstly, the run time of arithmetic operations is very much on the big digital image, I , and its two separated regions, Part1 and Part2. We are use the linear array (probability distribution) rather than I , for segmentation operation, and threshold values computation , and . Secondly, rather than we are create many binary matrices and apply the edge detector procedure for each region individually, then merge the resultant images into one. We are create one binary matrix according to

threshold values , and together, then apply the edge detector procedure one time. This modifications will reduce the run time of computations.

V. EXPERIMENTAL RESULTS

To demonstrate the efficiency of the proposed approach, the algorithm is tested over a number of different grayscale images and compared with traditional operators. The performance of the method is tested under noisy condition (Salt & Pepper noise) on test images. The images are corrupted by Salt & Pepper noise with 5%, 15% and 30% noise density before processing. The images detected by Canny, LOG, Sobel, Prewitt and the proposed method, respectively. All the concerned experiments were implemented on Intel® Core™ i3 2.10GHz with 4 GB RAM using MATLAB R2007b. As the algorithm has two main phases – global and local enhancement phase of the threshold values and detection phase, we present the results of implementation on these images separately.

The proposed scheme used the good characters of Kapur entropy, to calculate the global and local threshold values. Hence, we ensure that the proposed scheme done better than the traditional methods.

In order to validate the results, we run the Canny, LOG, Sobel and Prewitt methods and the proposed algorithm 10 times for each image with different sizes. As shown in Fig. 4. It has been observed that the proposed edge detector works effectively for different gray scale digital images as compare to the run time of Canny method.

Some selected results of edge detections for these test images using the classical methods and proposed scheme are shown in Fig.(6-7). From the results; it has again been observed that the performance of the proposed method works well as compare to the performance of the previous methods (with default parameters in MATLAB).

Fig. 5: Chart time for proposed method and classical methods with 512×512 pixel test images

0

0.5

1

1.5

2

ProposedSobelLOGCannyPrewitt

Avarage tim

e "Secon

d"

ALGORITHM 2: EDGE DETECTION

1. Input: A grayscale image I of size and , that has been calculated from algorithm 1.

2. Create a binary image: For all x, y,

if , then , 0 else , 1. 3. Create a mask w of order , in our case ( 3, 3) 4. Create an output image : For all x and y, Set

, , . 5. Checking for edge pixels:

Calculate: 1 /2 and 1 /2. For all 1 , … , , and 1 , … , , 0; For all , … , , and , … , , if ( , , ) then 1. if ( 6 ) then , 0 else , 1 .

6. Output: The edge detection image of I.



Original 0% noise

5% noise

15%noise

30% noise Canny method LOG method Sobel method Prewitt method Proposed method

Fig. 6: Performance of Proposed Edge Detector for Cameraman image with Various salt and pepper noise

Original 0%

5% noise

15%noise

30% noise Canny method LOG method Sobel method Prewitt method Proposed method

Fig. 7: Performance of Proposed Edge Detector for Blood cells image with Various salt and pepper noise



VI. CONCLUSION

An efficient approach using Kapur entropy for detection of edges in grayscale images is presented in this paper. The proposed method is compared with traditional edge detectors. On the basis of visual perception and edgel counts of edge maps of various grayscale images it is proved that our algorithm is able to detect highest edge pixels in images. The proposed method is decrease the computation time as possible as can with generate high quality of edge detection. Also it gives smooth and thin edges without distorting the shape of images. Another benefit comes from easy implementation of this method.

REFERENCES

[1] R. C. Gonzalez and R.E. Woods, "Digital Image Processing.", 3rd Edn., Prentice Hall, New Jersey, USA. ISBN: 9780131687288, 2008.

[2] F. Ulupinar and G. Medioni, “Refining Edges Detected by a LoG operator”, Computer vision, Graphics and Image Processing, 51, 1990, 275-298.

[3] J. Canny, "A Computational Approach to Edge Detect", IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol.PAMI-8, No.6, 1986, 679-698.

[4] Dong Hoon Lim, "Robust edge detection in noisy images", Computational Statistics & Data Analysis,50, 2006, pp. 803-812.

[5] Amiya Halder, etl., " Edge Detection: A statistical Approach", 3rd International Conference on Electronics Computer Technology (ICECT 2011).8-10April-2011 Kanyakumari, India

[6] Kaustubha Mendhurwar, etl., " Edge Detection in Noisy Images Using Independent Component Analysis", ISRN Signal Processing , Volume 2011, Article ID672353, 9 pages.

[7] Bijay Neupane, etl., " A New Image Edge Detection Method using Quality-based Clustering", Technical Report DNA#2012-01, April2012, Abu Dhabi, UAE.

[8] A. El-Zaart, "A Novel Method for Edge Detection Using 2 Dimensional Gamma Distribution", Journal of Computer Science 6 (2), 2010 , pp. 199-204.

[9] R. M. Haralick, “Digital Step Edges from Zero Crossing of Second Directional Derivatives”, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. PAMI-6, No1, Jan, 1984, 58-68.

[10] V. S. Nallwa and T. O. Binford, “On Detecting Edges”, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. PAMI-8, No.6, Nov, 1986, 699-714 .

[11] V. Torre and T. A. Poggio, “On Edge Detection”, IEEE Trans. on Pattern Analysis and Machine Intelligenc, Vol. PAMI-8, No.2, Mar. 1985, 147-163.

[12] D. J. Willians and M. Shan, “Edge Contours Using Multiple Scales”, Computer Vision, Graphics and Image Processing, 51, 1990, pp. 256-274.

[13] A. Goshtasby, “On Edge Focusing”, Image and Vision Computing, Vol. 12, No.4, May, 1994, 247-256 .

[14] A. Huertas and G. Medioni, “Detection of Intensity Changes with Subpixel Accuracy Using Laplacian-Gaussian Masks”, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. PAMI-8, No.5, Sep.1986, 651-664 .

[15] H. D. Cheng and Muyi Cu, “Mass Lesion Detection with a Fuzzy Neural Network”, J. Pattern Recognition, 37, pp.1189-1200, 2004.

[16] Jun Xu and Jinshan Tang, “Detection of Clustered Micro calcifications using an Improved Texture based Approach for Computer Aided Breast Cancer Diagnosis System”, J. CSI Communications, vol. 31, no. 10, 2008, pp. 17-20.

[17] M. Basu, "A Gaussian derivative model for edge enhancement.", Patt. Recog., 27:1451-1461, 1994.

[18] C. Kang, and W. Wang, "A novel edge detection method based on the maximizing objective function.", Pattern. Recog., 40, 2007, pp. 609-618.

[19] M. Roushdy, "Comparative Study of Edge Detection Algorithms Applying on the Grayscale Noisy Image Using Morphological Filter", GVIP, Special Issue on Edge Detection, 2007, pp. 51-59.

[20] B. Mitra, "Gaussian Based Edge Detection Methods- A Survey ". IEEE Trans. on Systems, Manand Cybernetics , 32, 2002, pp. 252-260.

[21] F. Luthon, M. Lievin and F. Faux, "On the use of entropy power for threshold selection." Int. J. Signal Proc., 84, 2004, pp. 1789-1804.

[22] M.Sonka, V.Hlavac, and R.Boyle, "Image Processing, Analysis, and Machine Vision" Second Edition, ITP, 1999.

[23] Shannon, C.E., “A mathematical Theory of Communication”, Int. J. Bell. Syst. Technical, vol.27, pp. 379-423, 1948.

[24] Baljit Singh and Amar Partap Singh, “Edge Detection in Gray Level Images based on the Shannon Entropy”, J. Computer Science, vol.4, no.3, 2008, pp.186-191.

[25] Landsberg, P.T. and Vedral, V. Distributions and channel capacities in generalized statistical mechanics. Physics Letters A, 247, (3), 1998, pp. 211-217.

[26] Alfréd Rényi , On measures of entropy and information, Pro ceeding of fourth Berkeley Symposium on Mathematical statistics and Probability, Vol. 1, 1961, pp. 547-561.

[27] M. P. de Albuquerque, I. A. Esquef , A.R. Gesualdi Mello, "Image Thresholding Using Tsallis Entropy." Pattern Recognition Letters 25, 2004, pp. 1059–1065.

[28] C. Tsallis, Non-extensive Statistical Mechanics and Themodynamics, Historical Background and Present Status. In: Abe, S. , Okamoto, Y. (eds.) Nonextensive Statistical Mechanics and Its Applications, Springer-Verlag Heidelberg, 2001.

[29] C. Tsallis, , Possible generalization of Boltzmann-Gibbs statistics, Journal of Statistical Physics, 52,, 1988, pp. 479-487.

[30] J. N. Kapur, H. K. Kesavan, Entropy optimisation Principles with Applications, Academic Press, Boston,1994.

[31] J. N. Kapur, Generalization entropy of order α and type β, The mathematical seminar, 4 , 1967, pp 78-84.


(IJCSIS) International Journal of Computer Science and Information Security,

Vol. 11, No. 12, December 2013

Influence of Stimuli Color and Comparison of SVM

and ANN classifier Models for BCI based

Applications using SSVEPs

Rajesh Singla Department of Instrumentation and Control

Engineering,

Dr. B. R. Ambedkar National Institute of Technology

Jalandhar, Punjab-144011, India

Arun Khosla

Department of Electronics and Communication

Engineering,

Dr. B. R. Ambedkar National Institute of Technology

Jalandhar, Punjab-144011, India

Rameshwar Jha Director General, IET Bhaddal,

Distt.- Ropar, Punjab-140108 ,India

Abstract—— In recent years, Brain Computer Interface (BCI)

systems based on Steady-State Visual Evoked Potential

(SSVEP) have received much attentions. In this study four

different flickering frequencies in low frequency region were

used to elicit the SSVEPs and were displayed on a Liquid

Crystal Display (LCD) monitor using LabVIEW. Four stimuli

colors, green, blue, red and violet were used in this study to

investigate the color influence in SSVEPs. The

Electroencephalogram (EEG) signals recorded from the

occipital region were segmented into 1 second window and

features were extracted by using Fast Fourier Transform

(FFT). This study tries to develop a classifier, which can

provide higher classification accuracy for multiclass SSVEP

data. Support Vector Machines (SVM) is a powerful approach

for classification and hence widely used in BCI applications.

One-Against-All (OAA), a popular strategy for multiclass SVM

is compared with Artificial Neural Network (ANN) models on

the basis of SSVEP classifier accuracies. Based on this study, it

is found that OAA based SVM classifier can provide a better

results than ANN. In color comparison SSVEP with violet

color showed higher accuracy than that with other stimuli .

Keywords- Steady-State Visual Evoked Potential; Brain

Computer Interface; Support Vector Machines; ANN.

I. INTRODUCTION

The Brain Computer Interface (BCI) system provides a direct communication channel between human brain and the computer without using brain‟s normal output pathways of peripheral nerves and muscles [1]. By acquiring and translating the brain signals that are modified according to the intentions, a BCI system can provide an alternative,

augmentative communication and control options for individuals with severe neuromuscular disorders, such as spinal cord injury, brain stem stroke and Amyotrophic Lateral Sclerosis (ALS).

Electroencephalography (EEG) is a non-invasive way of acquiring brain signals from the surface of human scalp, which is widely accepted due to its simple and safe approach. The brain activity commonly utilized by EEG based BCI systems are Event Related Potentials (ERPs), Slow Cortical Potentials (SCPs), P300 potentials, Steady-State Visual Evoked Potentials (SSVEPs) etc. Among them SSVEPs are attracted due to its advantages of requiring less or no training, high Information Transfer Rate (ITR) and ease of use [1, 2, 3].

SSVEPs are oscillatory electrical potential that are elicited in the brain when the person is visually focusing his/her attention on a stimulus that is flickering at frequency 6Hz or above [4]. These signals are strong in occipital region of the brain and are nearly sinusoidal waveform having the same fundamental frequency as the stimulus and including some of its harmonics. By matching the fundamental frequency of the SSVEP to one of the stimulus frequencies presented, it is possible to detect the target selected by the user. Considering the amplitudes of SSVEPs induced, the stimuli frequencies are categorized into three ranges, centered at 15 Hz low frequency, 31 Hz medium frequency and 41 Hz high frequency respectively [5].

There are many research groups that are designing SSVEP based BCI systems. Lalor et al. [6] developed the


mailto:[email protected]




control for an immersive 3D game using SSVEP signal. Muller and Pfurtscheller [7] used SSVEPs as the control mechanism for two-axis electrical hand prosthesis. Recently, Lee et al. [8] presented a BCI system based on SSVEP to control a small robotic car.

One of the main considerations during the development of a BCI system is to improve the classifiers accuracy, as that can affect the overall system accuracy and thus the ITR. In this research work, comparative study of Artificial Neural Network (ANN) and Support Vector Machine (SVM) have been carried out based on the classification accuracy of a multiclass SSVEP signal.

The retina of human eye contains rod and cone cells. The rod cells detect the amount of light and cone cells distinguish the color. There are three kinds of cone cells and are conventionally labeled as Short (S), Medium (M), and Long (L) cones according to the wavelengths of the peaks of their spectral sensitivities. S, M and L cone cells are therefore sensitive to blue (short-wavelength), green (medium-wavelength) and red (long-wavelength) light respectively. The brain combines the information from each cone cells to give different perceptions to different colors; as a result, the SSVEP strength elicited with different colors of the stimuli will different.

II. MATERIALS AND METHODS

A. Subject

Twenty right handed healthy subjects (17 males and 3 females, aged 22-27 years), with normal or corrected to normal vision participated in the experiment. All of them had normal color vision and not had any previous BCI experience. Prior starting, subjects were informed about the experimental procedure and required to sign a consent form. Table I shows the clinical characteristics of subjects.

CLINICAL CHARACTERISTICS OF SUBJECTS

S. No. Subject Age Education Status

1. Subject 1 22 B.Tech

2. Subject 2 24 M.Tech



















B. Stimuli

The template is used to format your paper and style the text. All margins, column widths, line spaces, and text fonts are prescribed; please do not alter them. You may note peculiarities. For example, the head margin in this template measures proportionately more than is customary. This measurement and others are deliberate, using specifications that anticipate your paper as one part of the entire proceedings, and not as an independent document. Please do not revise any of the current designations.

The RVS for eliciting SSVEP responses can be presented on a set of Light Emitting Diodes (LEDs) or on a Liquid Crystal Display (LCD) monitor [9]. In this study RVS displayed using LCD monitor due to the flexibility in changing the color of flickering bars, and were designed using LabVIEW software (National Instrument Inc., USA). Four colors: blue, green, red and violet were included in the experiment. Background color selected as black. Four frequencies 7, 9, 11 and 13 Hz, in the low frequency range were selected, as the refreshing rate of LCD monitor is 60 Hz [10] and the high amplitude SSVEPs are obtained at lower frequencies. The visual stimuli were square (4cm×4cm) in shape and were placed on four corners of the LCD screen. In order to select any particular stimuli the four visual stimuli were separated in pair of two each 7,11 and 9,13.Further in a interval of 2 sec if eye blink once then first pair was selected i.e 7,11 similarly if in that same interval if it blink twice then the next pair was selected i.e 9,13. Once a pair of stimuli was selected then again in next interval of 2sec if eye blink once then upper stimuli was selected and if it blink twice then the lower stimuli was selected in that pair of stimuli.

C. Experimental setup

The subjects were seated 60cm in front of the visual stimulator as shown in Fig.1. EEG signals were recorded using RMS EEG-32 Super Spec system (Recorders and Medicare System, India). The SSVEP potential recorded from occipital region using Ag/AgCl electrodes were amplified and connected to the adaptor box through head box. Adaptor box consist the circuitry for signal conditioning and further connected to the computer via USB port. This system can record 32 channels of EEG data. The electrodes were placed as per the international 10-20 system. The skin-electrode impedance was maintained below 5KΩ. The EEG signals were filtered by using a 3-50 Hz band pass filter and




a 50 Hz notch filter. Signals were sampled at 256 Hz and the sensitivity of the system was selected as 7.5µV/mm.

In training session the electrodes were placed at the O1, O2 and Oz regions of the scalp. The reference electrodes were placed on the right and left earlobes (A1 and A2) and ground electrode on Fpz. We first collected the SSVEP data for all the four frequencies with green color and then repeated the experiment for violet, blue and red color in another session. The interval between the sessions was 10 minutes. Initially the subjects were required to close their eyes for recording 2 minutes of baseline signal and then given 5 minutes to adapt to the flickering stimulus placed in front of them.

Figure 1. Experimental set up for SSVEP data acquisition (Courtesy-

Department of Instrumentation and Control Engineering, National Institute of Technology, Jalandhar)

During experiments, the subjects were directed to focus on a particular frequency for 5 second duration followed by 5 second rest period. During focusing the subjects were instructed to avoid eye movements or blinking. The event markers were used to indicate the starting and ending time of each frequency. In a single trial, each of the four frequencies was performed three times and the same procedure was repeated for another three trials. 5 minutes break was given in between each trial. The time for completing one session was about 30 minutes.

D. Feature Extraction

The frequency features of SSVEPs can easily extracted by using Fast Fourier Transform (FFT) [11]. The EEG signals recorded from Oz -A2 channel were digitized and segmented into 1 second time window in every 0.25seconds. MATLAB was used for developing the FFT program. Fig. 2 shows the amplitude spectra of SSVEP induced by 7 Hz stimulation. The coefficients at the fundamental and second harmonics of all the four target frequencies obtained from the amplitude spectra were considered as the feature vector for classification.

Figure 2. Amplitude spectra of SSVEP in response to 7 Hz, recorded from

Oz -A2 channel of subject 4.

E. Classification

ANN and SVM classifiers were implemented to classify the feature vectors and compared with respect to the classification accuracy. Multilayer ANN architecture consists of an input layer, a number of hidden layers and an output layer. Back propagation [12] is a supervised learning algorithm which can be used in multilayer ANN. This algorithm involves a forward propagation of input data through the network for calculating output values. Then the error obtained from the comparison between the output and target values are back propagated to adjust the weights of the neurons in each layer.

Two ANN models, Feed-forward Back propagation (FFBP) and Cascade-forward Back propagation (CFBP) were designed. In FFBP neurons are connected in feed forward fashion from the input layer to the output layer through the hidden layers according to back propagation algorithm. CFBP is similar to FFBP in using back propagation algorithm, with an exception that they have a weight connection from the input and every previous layer to the following layer and thus each layer neuron relates all previous layer neurons including input layer.

Modeling of the ANN was by using MATLAB neural network training tool. The input and output data were normalized in the range of [-1, +1]. Different combinations of internal parameters, such as number of hidden layers, number of neurons in each hidden layer, transfer function of hidden layers and output layer etc were tried. The input layer requires eight neurons by considering the first and second harmonics of each of the four frequencies. The output layer has four neurons corresponding to four frequencies. Gradient descent with momentum weight and bias learning function was used in both FFBP and CFBP models. Different variants of the back propagation algorithm were explored like Levenberg-Marquardt back propagation, Fletcher-Powell conjugate gradient back propagation and Bayesian regularization.

Performance of the ANN model was measured by Mean Square Error (MSE) function. The Cross Validation (CV) procedure [12] evaluates the training and learning of the NN model. The CV is executed at the end of training epoch and uses two independent data sets: the training set and the validation set for evaluating the training and learning errors [16].




The SVM technique introduced by Vapnik in [13] is basically a binary classifier which can discriminate between two classes by using an optimal hyper plane which maximize the margin between the two classes. Kernel functions provide a convenient method for mapping the data space into a high-dimension feature space without computing the non-linear transformation [14]. The common kernel functions are linear, quadratic, polynomial and radial basis function (RBF).

SVM training and classification was done by using MATLAB Bioinformatics toolbox. One-Against-All (OAA) method [15] was adopted for getting a multiclass SVM. The formulation of this mode states that a data point would be classified under a certain class if that class‟s SVM accepted it while rejected by all other classes SVMs. In this mode four binary SVMs were trained, each for one of the four frequencies. After training, there develop a structure having the details of the SVM like the number of support vectors, alpha, bias etc.

III. RESULTS AND DISCUSSIONS

The feature vector extracted using FFT were used for classification. There have four separate data sets each for four different stimuli colors. The training data set for each color consist 25 samples (5 samples for each of the four frequencies and 5 for rest signal) from each subject data i.e. a total of 500 samples in a complete set. The data were normalized in the range of [-1, +1]. After dozens of training sessions, an ANN network configuration having one hidden layer with 10 neurons was selected. Levenberg-Marquardt back propagation algorithm gave better results as compared to other training algorithms. For SSVEP classification, pure linear and tangent sigmoid functions were found better for hidden and output layer neurons respectively. Fig. 3 presents the MSE performance measures for FFBP and CFBP classifiers by using different color data. The CFBP algorithm converges at a faster rate than FFBP. The best validation performance of FFBP is 0.1214 and that of CFBP is 0.009213 at epoch 12. It is clear that the performance of CFBP is better than FFBP.

a. MSE performance measure of FFBP using violet colour data b. MSE performance measure of CFBP using violet colour

c. MSE performance measure of FFBP using green colour data d. MSE performance measure of CFBP using green colour data




e. MSE performance measure of FFBP using blue colour data f. MSE performance measure of CFBP using blue colour data

g. MSE performance measure of FFBP using red colour data h. MSE performance measure of CFBP using red colour data

Figure3. MSE performance measure of FFBP and CFBP during training using SSVEPs elicited by different stimuli

Individual SVMs were trained with different kernel

functions. Here Table II shows comparison of various kernel functions for violet color. Quadratic kernel provides an accuracy of 92.8% for 9 Hz signal. Higher accuracy for 11 Hz and 13 Hz were provided by linear kernel and is 88.80% and 100% respectively. Polynomial Kernel provided an accuracy of 100% for 7 Hz. All these selected kernels used less number of support vectors for particular classes.

Fig. 4 presents the regression plots for FFBP, CFBP and OAA-SVM classifiers during training using different data. The regression values obtained during training of violet data is 0.8446, 0.8773 and 0.9323 for FFBP, CFBP and SVM classifiers respectively. The regression values obtained during training of green data is 0.8224, 0.8536 and 0.9052 for FFBP, CFBP and SVM classifiers respectively. The regression values obtained during training of blue data is

0.82651, 0.8502 and 0.9010 for FFBP, CFBP and SVM classifiers respectively. The regression values obtained during training of red data is 0.8129, 0.8383 and 0.8916 for FFBP, CFBP and SVM classifiers respectively. This proves the superior performance of OAA-SVM over FFBP and CFBP for SSVEP classification.

The designed classifier models were tested using the data

sets obtained from twenty subjects with 10 samples from

each subject in each class(i.e., SSVEP features for 7Hz,

9Hz, 11Hz, 13Hz and rest signals) i.e. total 50 samples from

one subject and total of 1000 samples for each stimuli color.

The accuracies obtained during testing of the data sets using

the same configurations of FFBP, CFBP and OAA-SVM

classifiers are presented in Table III.




(VIOLET COLOR)

Kernel

Function

7 Hz 9 Hz 11 Hz 13 Hz

Accuracy

(%)

Support Vectors

Accuracy

(%)

Support

Vectors

Accuracy

(%)

Support

Vectors

Accuracy

(%)

Support

Vectors

Linear 95 14 79 25 88.8 16 100 18

Quadratic 97.8 23 92.8 25 83.4 20 37.8 21

Polynomial

(order 3)

99 23 71.3 37 81.8 22 67.3 22

Polynomial

(Order 4)

100 38 89.1 37 87.8 34 19.8 25

Radial

Basis Function

35 118 37.2 117 68.3 119 38.6 116

Compared with FFBP and CFBP the OAA-SVM gave higher accuracy for all colors. Accuracy of CFBP is higher than the accuracy of FFBP but lower than that of OAA-SVM.Experimental result suggested that, for a multiclass SSVEP data OAA-SVM can give better classification accuracy than that of FFBP and CFBP models.

Test data result of violet stimuli shows higher accuracy than that of other color with all three classifiers. Violet color has an average accuracy of 93.23% with OAA-SVM classifier and is higher than the accuracy of green

with the same classifier. The reason for this may be related to the principle of perception of light and color sensitivity of human eyes. As mentioned, the other color can only elicit medium cone cells. Violet color, a combination of blue and red can elicit the cones responsible for blue and red, i.e. both short and long cones. As a result, with violet color more intense SSVEP is appeared in visual cortex (occipital lobe) of the brain compared to other color.

FFBP - Regression plot for violet stimuli CFBP - Regression plot for violet stimuli




OAA-SVM -Regression plot for violet stimuli FFBP - Regression plot for green stimuli

O

CFBP - Regression plot for green stimuli OAA-SVM - Regression plot for green stimuli

FFBP - Regression plot for blue stimuli CFBP - Regression plot for blue stimuli




OAA-SVM - Regression plot for blue stimuli FFBP - Regression plot for red stimuli

CFBP-Regression plot for red stimuli OAA-SVM-Regression plot for red stimuli

Figure 4. Comparison of regression plots of FFBP, CFBP and OAA-SVM models obtained during SSVEP data classification

TABLE III COMPARATIVE RESULTS OF TESTING ACCURACY OF SSVEPS ELICITED BY GREEN(G), RED(R) , BLUE(B) AND VIOLET(V) COLOR STIMULI FOR 20 SUBJECTS BASED ON ANN AND SVM CLASSIFIERS

Stimuli color FFBP CFBP OAA-SVM

Blue 82.65 85.02 90.10

Green 82.24 85.36 90.52

Red 81.29 83.83 89.16

Violet 84.46 87.73 93.23




IV. CONCLUSIONS

In this research three classifier models (FFBP, CFBP and OAA-SVM) were constructed for SSVEP data classification. The motivation of this work is to improve the accuracy of SSVEP based BCI system by improving the classification accuracy. EEG signals were recorded by using RMS EEG-32 Super Spec system and SSVEP features extracted using FFT. SSVEPs were elicited using four different frequencies. Four different stimuli color, green, red, blue, and violet were compared to get better performance. The amplitudes of first and second harmonics of SSVEP data were successfully used as the feature vector to train the classifier models. The experimental result shows that OAA-SVM yields superior classification accuracy compared against FFBP and CFBP for SSVEP data. The result also showed that SSVEPs with violet stimuli is better than that with other stimuli.

The future work may include the development of a SSVEP based BCI application system that can provide higher accuracy by using OAA-SVM classifier.

ACKNOWLEDGMENT

The authors would like to thank the subjects who

participated in the EEG recording session.

REFERENCES

[1] J. R. Wolpaw, N. Birbaumer, W. J. Heetderks, D. J.

McFarland, P. H. Peckham, G. Schalk, E. Donchin, L. A.

Quatrano, C. J. Robinson, and T. M. Vaughan, “Brain Computer Interface technology: - a review of the first

international meeting,” IEEE Trans. Rehab. Eng. vol. 8,

pp. 164-173, June 2000. [2] R. Singla, B. Chambayil, A. Khosla, J. Santosh

“Comparison of SVM and ANN for classification of eye

events in EEG,” Journal of Biomedical Sciences and engineering (JBISE), Vol 4, No 2, January 2011, pg (62-

69). Scientific Research Publishing, USA

[3] T. W. Berger, J. K. Chapin, G. A. Gerhardt, et. al., “International assessment of research and development in

brain-computer interfaces: report of the world technology

evaluation center,” Berlin: Springer, 2007. [4] M. Cheng, X.R. Gao, S.K. Gao, and D. xu, “Design and

implementation of a brain computer interface with high

transfer rates,” IEEE Trans Biomed Eng., 2002, vol. 49. No. 10, pp. 1181-1186

[5] Y. J. Wang, R. P. Wang, X. R. Gao, B. Hong, and S. K.

Gao, “A Practical VEP-based brain-computer interface,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 14, no. 2, pp.

234-240, June 2006.

[6] E. C. Lalor, S. P. Kelly, C. Finucane, R. Burke, R. Smith,

R. B. Reilly, and G. McDarby, „„Steady-state VEP-based

brain-computer interface control in an immersive 3D gaming environment,” EURASIP J. Appl. Signal Process.,

vol. 2005, no. 19, pp. 3156–3164, 2005.

[7] G. R. Muller-Putz and G. Pfurtscheller, “Control of an electrical prosthesis with an SSVEP-based BCI,” IEEE

Trans. Biomed. Eng., vol. 55, no. 1, pp. 361–364, 2008.

[8] P. L. Lee, H. C. Chang, T. Y. Hsieh et.al., “A brain wave

actuated small robot car using ensemble empiricalmode

decomposition based approach,” IEEE Trans. Sys. Man

and cyber.part A: Systems and humans, vol. 42, no. 5, pp

1053-1064, sept. 2012. [9] D. H. Zhu, J. Bieger, G. G. Molina, and R. M. Aarts, “A

Survey of Stimulation Methods Used in SSVEP-Based

BCI system,” Comput. Intell. Neurosci., pp. 702357, 2010. [10] Y. Wang, Y. –T. Wang, and T. –P. Jung, “Visual stimulus

design for high-rate SSVEP BCI,” Electron. Lett., vol 46,

No. 15, pp. 1057-1058, 2010. [11] G. R. Muller-Putz, R. Scherer, C. Brauneis, and G.

Pfurtscheller, “Steady-State Visual Evoked Potential

(SSVEP)-based Communication: Impact of Harmonic Frequency Components,” J. Neural Eng., vol. 2, no. 4, pp.

123-130,2005.

[12] S. Haykin, “Neural Networks:A Comprehensive Foundation,”1998, Prentice Hall.

[13] V. Vapnik., “Statistical Learning Theory,” John Wiley and

Sons, Chichester, 1998. [14] X. Song-yun, W. Peng-wei, Z. Hai-jun, Z. Hai-tao,

“Research on the classification of Brain function based on

SVM,” The 2nd International Conference on Bioinformatics and Biomedical Engineering, ICBBE 2008,

p1931 – p1934, 2008.

[15] B. Chambayil, R. Singla, R. Jha, “Virtual keyboard BCI using eyeblinks in EEG,” IEEE 6th int. conf. on wireless

and mobile computing, networking and communication, pp 466-470, 2010

[16] Vijay Khare, Jayashree Santosh, Sneh Anand and Manvir

Bhatia “Classification of five mental tasks based on two methods of neural network”. International Journal of

Computer science and information security ,pp 86-92,

Vol8 No.3, 2010 AUTHORS PROFILE

Rajesh Singla was born in Punjab , India in 1975. He obtained B.E Degree from Thapar University in 1997, M.Tech degree froom IIT-Roorkee in 2006. Currentely he is pursuing Ph.D degree from National Institute of Technology Jalandhar, Punjab, India. His area of interest is Brain Computer Interface, Rehabilation Engineering, and Process Control.

He is working as an Associate Professor in National Institute of Technology Jalandhar, India.

Dr Arun Khosla was born in Punjab, India. He received the BE degree from Thapar University, India, M.Tech from NIT Kurukshetra, Ph.D. from Kurukshetra University, India. His research areas include Artificial Intelligence and Bio-Medical Instrumentation.

He is working as an Associate Professor in the Department of Electronics and Communication Engineering, NIT Jalandhar. He is also Head of the Department since 2010.

Dr R. Jha was born in Bihar, India in 1945. He received his

BSc (Engineering Electrical)from Bhagalpur University in 1965, his MTech From IIT-Delhi in 1970 and his PhD degree from

Punjab University-Chandigarh, India, in 1980. He worked as a

Lecturer in the Johrat Engineering Collage-Johrat, (1965–1972) and as an Assistant Professor at the Punjab Engineering College

Chandigarh, (1972–1986). In 1986, he joined REC-Hamirpur, as

a Professor and rose to be its Principal in 1990. He joined NIT-Jalandhar as professor and Head of the Instrumentation &

Control Engineering Department (1994–2010). His area of

interest is computer-aided analysis and design of control systems and fuzzy systems.

He is working as the Director- Principal of the Institute of

Engineering and Technology-Bhaddal, India.


(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

Comparative Study of Person Identification System with Facial Images Using PCA and KPCA

Computing Techniques

Md. Kamal Uddin, Abul Kalam Azad, Md.Amran Hossen Bhuiyan Department of Computer Science & Telecommunication Engineering, Noakhali Science & Technology University

Noakhali-3814, Bangladesh

Abstract— Face recognition is one of the most successful areas of research in computer vision for the application of image analysis and understanding. It has received a considerable attention in recent years both from the industry and the research community. But face recognition is susceptible to variations in pose, light intensity, expression, etc. In this paper, a comparative study of linear (PCA) and nonlinear (KPCA) based approaches for person identification has been explored. The Principal Component Analysis (PCA) is one of the most well-recognized feature extraction tools used in face recognition. The Kernel Principal Component analysis (KPCA) was proposed as a nonlinear extension of a PCA. The basic idea of KPCA is to maps the input space into a feature space via nonlinear mapping and then computes the principal components in that feature space. In this paper, facial images have been classified using Euclidean distance and performance has been analysed for both feature extraction tools. Keywords—Face recognition; Eigenface; Principal component analysis; Kernel principal component analysis.

I. INTORDUCTION

Modern civilization heavily depends on person authentication for several purposes. Face recognition has always a major focus of research because of its non-invasive nature and because it is people’s primary method of person identification. The identification of a person interacting with computers represents an important task for automatic systems in the area of information retrieval, automatic banking and control of access to security areas and so on. Here, a tiny effort has been carried out to develop the person identification systems with facial image by using Principal Component Analysis (PCA) and Kernel Principal Component Analysis (KPCA) computing techniques. And finally the performance has been compared with both computing techniques.

II. RELATED WORKS Face recognition is an active area of research with applications ranging from static, controlled mug-shot verification to dynamic, uncontrolled face identification in a cluttered

background [1]. In the context of personal identification, face recognition usually refers to static, controlled full-frontal portrait recognition [10]. By static, it means that the facial portraits used by the face recognition system are still facial images (intensity or range). By controlled, it means that the type of background, illumination, resolution of the acquisition devices, and the distance between the acquisition devices and faces, etc. are essentially fixed during the image acquisition process. Obviously, in such a controlled situation, the segmentation task is relatively simple and the intra-class variations are small.

Over the past three decades, a substantial amount of research effort has been devoted to face recognition [1], [10]. In the1970s, face recognition was mainly based on measured facial attributes such as eyes, eyebrows, nose, and lips, chin shape, etc. [1]. Due to the lack of computational resources and brittleness of feature extraction algorithms, only a very limited number of tests were conducted and the recognition performance of face recognition systems was far from desirable [1]. After the dormant 1980s, there was resurgence in face recognition research in the early 1990s. In addition to continuing efforts on attribute based techniques [11], a number of new face recognition techniques were proposed, including:

Principal Component Analysis(PCA) [1], [11], [12] Linear Discriminant Analysis(LDA) [13] A variety of neural network based techniques [14] Kernel Principal Component Analysis(KPCA)[15]

III. FACE FEATURE EXTRACTION PCA is powerful technique for extracting a structure from potentially high-dimensional data sets, which corresponds to extracting the eigenvectors that are associated with the largest eigenvalues from the input distribution. For faster computation of the eigenvectors and eigenvalues, singular value decomposition (SVD) is used. This eigenvector analysis has already been widely used in face processing [1], [2]. A kernel



PCA, proposed as a nonlinear extension of a PCA [3]–[5] computes the principal components in a high-dimensional feature space, which is nonlinearly related to the input space. A kernel PCA is based on the principle that since a PCA in feature space can be formulated in terms of the dot products in feature space, this same formulation can also be performed using kernel functions (the dot product of two data in ) without explicitly working in feature space. In this section two methods which are used in this study are explained briefly.

A. Principal Component Analysis(PCA)

A 2-D facial image can be represented as 1-D vector by concatenating each row (or column) into a long thin vector. Let’s suppose, an M vectors of size N (= rows of image × columns of image) representing a set of sampled images. 푝 ’s represents the pixel values.

푥 = [푝 … … …푝 ] , 푖 = 1, … ,푀 (1)

The images are mean centred by subtracting the mean image

from each image vector. Let m represent the mean image.

푚 = ∑ 푥 (2)

And let 푤 be defined as mean centred image

푤 = 푥 −푚 (3)

The primary goal is to find a set of 푒 ’s which have the largest possible projection onto each of the 푤 ’s. This work wish to find a set of M orthonormal vectors 푒 for which the quantity

휆 = ∑ (푒 푤 ) (4)

is maximized with the orthonormality constraint

푒 푒 = 훿 (5) It has been shown that the 푒 ’s and 휆 ’s are given by the eigenvectors and eigenvalues of the covariance matrix

퐶 = 푊푊 (6)

Where W is a matrix composed of the column vectors 푤 placed side by side. The size of C is N × N which could be enormous. For example, images of size 64 × 64 create the covariance matrix of size 4096×4096. It is not practical to solve for the eigenvectors of C directly. A common theorem in linear algebra states that the vectors 푒 and scalars 휆 can be obtained by solving for the eigenvectors and eigenvalues of the M×M matrix 푊 푊.

Let 푑 and 휇 be the eigenvectors and eigenvalues of 푊 푊, respectively.

푊 푊푑 = 휇 푑 (7)

By multiplying left to both sides by W

푊푊 (푊푑 ) = 휇 (푊푑 ) (8) Which means that the first M – 1 eigenvectors 푒 and eigenvalues 휆 of 푊푊 are given by 푊푑 and 휇 , respectively. 푊푑 needs to be normalized in order to be equal to 푒 . Since only sum up of a finite number of image vectors is used, M, the rank of the covariance matrix cannot exceed M – 1 (The –1 come from the subtraction of the mean vector m). The eigenvectors corresponding to nonzero eigenvalues of the covariance matrix produce an orthonormal basis for the subspace within which most image data can be represented with a small amount of error. The eigenvectors are sorted from high to low according to their corresponding eigenvalues. The eigenvector associated with the largest eigenvalue is one that reflects the greatest variance in the image. That is, the smallest eigenvalue is associated with the eigenvector that finds the least variance. They decrease in exponential fashion, meaning that the roughly 90% of the total variance is contained in the first 5% to 10% of the dimensions. A facial image can be projected onto 푀 (≪푀) dimensions by computing

훺 = [푣 푣 … 푣 ] (9) Where 푣 = 푒 푤 . 푣 is the 푖 coordinate of the facial image in the new space, which came to be the principal component. The vectors 푒 are also images, so called, eigenimages, or eigenfaces in our case, which was first named by [6]. They can be viewed as images and indeed look like faces.

Fig 1. Eigenfaces for the example image set



So, 훺 describes the contribution of each eigenface in representing the facial image by treating the eigenfaces as a basis set for facial images. The simplest method for determining which face class provides the best description of an input facial image is to find the face class 푘 that minimizes the Euclidean distance 휖 = ‖(훺 −훺 )‖ (10) Where 훺 is a vector describing the kth face class. If 휖 is less than some predefined threshold 휃 , a face is classified as belonging to the class k.

B. Kernel Principal Component Analysis The basic idea of kernel PCA is to first map the input data into x a feature space F via a nonlinear mapping Φ and then perform a linear PCA in F. Assuming that the mapped data are centred, i.e., ∑ 훷(푥 ) = 0, where M is the number of input data (the centring method in F can be found in [7] and [8]), kernel PCA diagnoses the estimate of the covariance matrix of the mapped data 훷(푥 ), defined as

퐶 = ∑ 훷(푥 )훷(푥 ) (11) To do this, the eigenvalue equation 휆푣 = 퐶푣 must be solved for eigenvalues 휆 ≥ 0 and eigenvector푣 ∈ 퐹0. As퐶푣 =1푀∑ (훷(푥 ) ∙ 푣)훷(푥 ), all solutions v with 휆 ≠ 0 lie

within the span of 훷(푥 ), … ,훷(푥 ), i.e., the coefficients 훼 (푖 = 1, … ,푀) exist such that

푣 = ∑ 훼 훷(푥 ) (12) Then the following set of equations can be considered:

휆(훷(푥 ) ∙ 푣) = (훷(푥 ) ∙ 퐶푣) for all 푖 = 1, … ,푀 (13) The substitution of (11) and (12) into (13) and the definition of an M × M matrix K by 퐾 ≡ 푘 푥 ,푥 = (훷(푥 ) ∙ 훷 푥 ) produces an eigenvalue problem which can be expressed in terms of the dot products of two mappings Solve 푀휆훼 = 퐾훼 For nonzero eigenvalues 휆 and eigenvectors 훼 =(훼 , … ,훼 ) subject to the normalization condition 휆 (훼 ∙훼 ) = 1. For the purpose of principal component extraction, the projections of x are computed onto the eigenvectors 푣 in F. For face feature extraction using kernel PCA which involves three layers with entirely different roles. The input layer is made up of source nodes that connect the kernel PCA to its environment. Its activation comes from the gray level values of the face image. The hidden layer applies a nonlinear mapping Φ from the input space to the feature space F, where

the inner products are computed. These two operations are in practice performed in one single step using the kernel k. The outputs are then linearly combined using weights 훼 resulting in an lth nonlinear principal component corresponding to Φ. Thereafter, the first q principal components (assuming that the eigenvectors are sorted in a descending order of their eigenvalue size) constitute the q-dimensional feature vector for a face pattern. By selecting the proper kernels, various mappings, Φ, can be indirectly induced. One of these mappings can be achieved by taking the d-order correlations between the entries, 푥 , of the input vector x. Since x represents a face pattern with 푥 as a pixel value, a PCA in F computes the dth order correlations of the input pixels, and more precisely the most important q of the dth order cumulants. Note that these features cannot be extracted by simply computing all the correlations and performing a PCA on such pre-processed data, since the required computation is prohibitive when is not small (d ˃ 2): for N dimensional input patterns, the dimensionality of the feature space F is (N+d-1)!/d!(N-1). However, this is facilitated by the introduction of a polynomial kernel, as a polynomial kernel with degree 푑(푘(푥, 푦) = (푥 ∙ 푦) ) corresponds to the dot product of two monomial mappings, 훷 [7], [9].

훷 (푥) ∙ 훷 (푦) = 푥 ∙. ..∙ 푥 ∙ 푦 ∙… ∙ 푦,…,

= 푥 ∙ 푦 = (푥 ∙ 푦)

IV. FACE RECOGNITION

PCA and kernel PCA compute the basis of a space which is represented by its training vectors. These basis vectors, actually eigenvectors, computed by PCA and KPCA are in the direction of the largest variance of the training vectors, earlier considered them as eigenfaces. Each eigenface can be viewed as a feature. When a particular face is projected onto the face space, its vector into the face space describes the importance of each of those features in the face. The face is expressed in the face space by its eigenface coefficients (or weights). It is possible to handle a large input vector, facial image, only by taking its small weight vector in the face space. This means that it is possible to reconstruct the original face with some error, since the dimensionality of the image space is much larger than that of face space. Each face in the training set is transformed into the face space and its components are stored in memory. The face space has to be populated with these known faces. An input face is given to the system, and then it is projected onto the face space. The system computes its distance from all the stored faces.



However, two issues should be carefully considered: 1. What if the image presented to the system is not a

face? 2. What if the face presented to the system has not

already learned, i.e., not stored as a known face? The first defect is easily avoided since the first eigenface is a good face filter which can test whether each image is highly correlated with itself. The images with a low correlation can be rejected. Or these two issues are altogether addressed by categorizing following four different regions:

1. Near face space and near stored face => known faces 2. Near face space but not near a known face =>

unknown faces 3. Distant from face space and near a face class => non-

faces 4. Distant from face space and not near a known class

=> non-faces Since a face is well represented by the face space, its reconstruction should be similar to the original; hence the reconstruction error will be small. Non-face images will have a large reconstruction error which is larger than some threshold 휃 . The distance ∈ determines whether the input face is near a known face.

V. EXPERIMENTAL RESULTS To analyse the performance of both methods, experiments were performed using AT&T face database. This database contains ten different images of each of 40 different persons. For all persons, the images were taken at different times, varying the lighting, facial expressions (open/closed eyes, smiling/not smiling) and facial details (glasses/ no glasses). The following three experiments have been done where the results of both identification systems are comparatively discussed:

A. First Experiment: In this experiment, a subset of the database was taken, which contains only 12 person’s images, each person has 10 distinct images, has been performed to ensure how well the both identification systems can identify each individual’s face. Here, 10 persons were selected as training set and other 2 persons were selected as test set. After performing this experiment, both systems recognized the faces as unknown. In this case accuracy of both systems was 100%.

B. Second Experiment

In this experiment face database, which contains 40 person’s images, each person has 10 different images, so total number of samples is 400. These samples were selected as training set.

After this, 5 images were randomly chosen from each person and constructed different number of test set. The accuracy of both identification systems was 100%.

C. Third Experiment

In this experiment, each set of 10 images for a person was randomly portioned into a training subset of 8 images and remaining 2 images were considered as test set. Since, there were 40 persons, so total number of samples was 320. These samples were selected as training set. From other images, different numbers of test set were constructed. Recognition rate using PCA and KPCA computing techniques is shown in Figure 2. Number of recognized samples and false recognized sample for both computing techniques is shown in Figure 3 and 4.

Fig 2. Recognition accuracy using PCA and KPCA

Fig 3. Number of false recognized samples



Fig 4. Number of recognized samples

VI. CONCLUSION The aim of this paper was to present a comparative study of person identification system using linear (PCA) and nonlinear (KPCA) based approaches. The basic of this comparative study has considered on eigenspace and kernel eigenspace methods. In this case, two different projection methods (PCA and KPCA) and one similarity measure method (Euclidean distance) were considered. After analysing the comparative study between linear and nonlinear techniques, it has been evident that although kernels enable us to work in a high dimension feature space, they do not always ensure better performance. Moreover, the choice of optimal kernel is not always obvious.

ACKNOWLEDGMENT All the authors wish to acknowledge AT&T Bell

Laboratories face database for helping by providing all the images that had been used to implement this work properly. Without their prior assistance, this work would be impossible to accomplish.

REFERENCES [1]. M. Turk and A. Pentland, “Eigenfaces for recognition,” J. Cogn.

Neurosci., vol. 3, no. 1, pp. 71–86, 1991. [2]. J. Zhang, Y. Yan, and M. Lades, “Face recognition: Eigenface, elastic

matching, and neural nets,” Proc. IEEE, vol. 85, pp. 1423–1435, Sept. 1997.

[3]. B. Schölkopf, A. Smola, and K. Müller, “Non-linear component analysis as a kernel eigenvalue problem,” Neural Comput., vol. 10, pp. 1299–1319, 1998.

[4]. K. Müller, S. Mika, G. Rätsch, K. Tsuda, and B. Schölkopf, “An introduction to kernel-based learning algorithms,” IEEE Trans. Neural Networks, vol. 12, pp. 181–201, Mar. 2001.

[5]. Schölkopf, A. Smola, and K. Müller, “Kernel principal component analysis,” in Advances in Kernel Methods-Support Vector Learning, B. Schölkopf, C. Burges, and A. Smola, Eds. Cambridge, MA: MIT Press, 1999, pp. 327–352.

[6]. M.A. Turk and A.P. Pentland, “Face Recognition Using Eigenfaces”, IEEE Conf. on Computer Vision and Pattern Recognition, pp. 586-591, 1991.

[7]. Schölkopf, A. Smola, and K. Müller, “Non-linear componentanalysis as a kernel eigenvalue problem,” NeuralComput., vol. 10, pp.1299–1319, 1998.

[8]. Schölkopf, A. Smola, and K. Müller, “Kernel principal component analysis,” in Advances in Kernel Methods-Support Vector Learning, B. Schölkopf, C. Burges, and A. Smola, Eds. Cambridge, MA: MIT Press, 1999, pp. 327–352.

[9]. K. Müller, S. Mika, G. Rätsch, K. Tsuda, and B. Schölkopf, “An introduction to kernel-based learning algorithms,” IEEE Trans. Neural Networks, vol. 12, pp. 181–201, Mar. 2001.

[10]. R. Chellappa, C. Wilson, and A. Sirohey, “Human and Machine Recognition of Faces: A Survey, “Proc. IEEE, vol. 83, no.5, pp. 705-740, 1995.

[11]. M. Kirby and L. Sirovich, “Application of the Karhunen-Loeve Procedure for the Characterization of Human Faces,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol.12, no.1,pp 103-108, Jan. 1990.

[12]. L. Sirovich and M. Kirby, “Low Dimensional Procedure for Characterization of Human Faces,” J. Optical Soc. Am. Vol. 4, no. 3, pp. 519-524, 1987.

[13]. D.L. Swets and J. Weng, “Using Discriminant Eigenfeatures for Image Retrieval,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 18, no. 8, pp. 831-836, Aug. 1996.

[14]. Valentin, H. Abdi, A.J.O’Toole, and G. Cottrell, “Connectionist Models of Face Processing: A Survey,” Pattern Recognition, vol.27, no.9,pp. 1209-1230, 1994.

[15]. Kwang In Kim, Keechul Jung, and Hang Joon Kim. Face recognition using kernel principal component analysis. Signal Processing Letters, IEEE, 9, Feb. 2002.



COLOR IMAGE ENHANCEMENT OF FACE IMAGES WITH DIRECTIONAL FILTERING APPROACH USING BAYER’S PATTERN

ARRAY Dr. S. Pannirselvam Research Supervisor & Head Department of Computer Science,

Erode Arts & Science College (Autonomous), Erode, Tamil Nadu, India .

S. Prasath Ph.D (Research Scholar) Department of Computer Science,

Erode Arts & Science College (Autonomous), Erode, Tamil Nadu, India .

ABSTRACT - Today, image processing penetrates into various fields, but till it is struggling in quality issues. Hence, image enhancement came into existence as an essential task for all kinds of image processings. Various methods are been presented for color image enhancement, especially for face image. In this paper various filters are used for face image enhancement. In order to improve of the image quality directional filtering approach using Bayer’s pattern are has been applied. In this method the color image are get decomposed into three color component array, then the Bayer’s pattern array is applied to enhance those color component and interpolate the three colors into a single RGB color image. The experimental result shows that this method provides better enhancement in term of quality when compared with the existing methods such as Bilinear Method, Gaussian Filter and Vector Median Filter. The peak Signal Noise Ratio (PSNR) and Mean Square Error (MSE) are been used for similarity measures. Keywords- VMF, GF, BM, PBPM, RGB, YbCr , PSNR, MSE

1. INTRODUCTION In the computer era there is a rapid growth in the field

of information technology and the security system was suffering from various issues. Today, criminals have been entered into the field of information technology called cyber crime. Lot of security systems has emerged to solve the various security issues such as password, username, secret codes, but failed due to cyber attacks. In order to overcome such security issues the biometric system has been emerged with various features such as face recognition, fingerprints recognition, gait, palm print, voice, signatures etc.

Every human being can identify a faces in a scene with no effort, with an automated system such objectives are the very challenging one due to various factors which affects the quality of the image. Hence, face recognition system has been used to verify the identity of an individual. It can be accomplished by matching process using various methods and features such as geometric, statistical, low-level features which are derived from face images. The demosaicking process plays a crucial role in the image enhancement with good quality. Naive, single channel, demosaicking schemes such as nearest neighbor replication, bilinear interpolation and cubic spline interpolation are usually provides less image quality. In this directional filtering approach using Bayer’s pattern array method the image are been decomposed. After interpolating the green channel, the algorithms are used to estimate the red and blue components under the assumption that the color of neighboring pixels is

similar. Minimizing differences in the RGB ratios between neighboring pixels usually does the estimation. Other algorithms interpolate color differences and not color ratios. They are able to compute the gradient for the green channel correctly (or for the red and blue channels). Presently digital cameras carry out color demosaicking process prior to compression, apparently due to the considerations of easy user interface and device compatibility. Color demosaicking triples the amount of raw data by generating R, G and B bands via color interpolation. Demosaicking process is used to reconstruct the R, G and B bands. This relieves the camera from the tasks of color demosaicking and color decorrelation and reduces the amount of input data to compression codec. The new workflow can potentially reduce on-camera computing power and input/output bandwidth. More importantly, the new design allows lossless or near-lossless compression of raw mosaic data, which is the main theme of the preprocessing. In recent research in color demosaicking indicates that more sophisticated color demosaicking algorithms than those implemented on camera, provided that original mosaic data are available, can obtain superior image quality. Furthermore, other image and video applications, such as super-resolution imaging and motion analysis should also benefits from lossless compression of color mosaic data, in which even sub pixel precision is much desired. The noise is characterized by its pattern and its probabilistic characteristics. There is a wide variety of noise types such as Gaussian noise, salt and pepper noise, poison noise, impulse noise, speckle noise. 2. RELATED WORK

An adaptive iterative histogram matching (AIHM) algorithm [1] for chromosome contrast enhancement especially in banding patterns. The reference histogram, with which the initial image needs to be matched, is created based on some processes on the initial image histogram.

An image enhancement algorithm [2] of video analysis and the CI value is used as the evaluation function which can provide a reference to the degree of enhancement. The video image enhancement algorithm based on the point analysis method of multi-dimensional biomimetic informatics and it works well based on the point analysis method of multi-dimensional biomimetic algorithm.



A stochastic resonance based technique [3] introduced for enhancement of very low-contrast images. In this technique an expression for optimum threshold has been derived. Gaussian noise of increasing standard deviation has been added iteratively to the low-contrast image until the quality of enhanced image reaches the maximum.

A fusion based approach [4] on Multi Scale Retinex with Color Restoration (MSRCR) that would provides better image enhancement. Lower dynamic range of a camera as compared to human visual system causes images taken to be extremely dependent on illuminant conditions. MSRCR algorithm enhances images taken under a wide range of nonlinear illumination conditions to the level that a user would have perceived it in real time.

F. Deepak Ghimire et al., [5] developed a method for enhancing the color images based on nonlinear transfer function and pixel neighborhood by preserving details. In this method, the image enhancement is applied only on the V (luminance value) component of the HSV color image and H and S component are kept unchanged to prevent the degradation of color balance between HSV components. Finally, original H and S component image and enhanced V component image are converted back to RGB image.

A multi-scale enhancement algorithm [6] in which they utilize LIP model they consider the characteristics of the human visual system. Then a new measure of enhancement based on just noticeable difference model of human visual system is used for evaluating the performance of the enhancement technique.

A content aware algorithm [7] that enhances dark images, sharpens edges, reveals details in textured regions and preserves the smoothness of flat regions. This algorithm produces an ad hoc transformation for each image, adapting the mapping functions to each image characteristic to produce the maximum enhancement. To analyze the contrast of the image in the boundary, textured regions they group the information with common characteristics.

A hybrid algorithm [8] is used to enhance the image and it uses the Gauss filter processing to enhance image details in the frequency domain and smoothens the contours of the image by the top-hat and bot-hat transforms in spatial domain.

A Bayer array [9] consists of alternating rows of red-green and green-blue filters. The bayer array contains twice as many green as red or blue sensors. Each primary color does not receive an equal fraction of the total area because the human eye is more sensitive to green light than both red and blue light. Redundancy with green pixels produces an image which appears less noisy and has finer detail than could be accomplished and when each color were treated equally.

Simple interpolation algorithm [10] is used for multivariate interpolation on a uniform grid, using relatively straightforward mathematical operations using only nearby instances of the same color component. In the simplest bilinear interpolation method the red value of a non-red pixel is computed as the average of the adjacent red pixels and similar for blue and green.

An adaptive [11] for algorithm estimates the depending on features of the area surrounding the pixel of interest. Variable Number of Gradients interpolation computes gradients near the pixel of interest and uses the lower gradients to make an estimate. Pixel grouping uses assumptions about natural scenery in making estimates. It has fewer color artifacts on natural images than the variable number of gradients method.

An adaptive homogeneity-directed [13] interpolation selects the direction of interpolation so as to maximize a homogeneity metric, thus typically minimizing color artifacts. Assuming the laws of colorimetry, two pixels sharing the same hue, but differing in intensities, will have the same R/G/B ratio [12]. This assumption is true in the case of digital sensors which have a nearly linear response to light and since all of the color enhancement is done only after the demosaicing is completed.

The second criterion is response to the Harris corner detection filter [14]. In natural images edges are sparse and corners are much sparser [15]. Since assume that highly detailed regions will contain many edges, to grade the demosaicking results according to the response to a corner detection filter alone. Due to zippering, erroneous demosaicking very often yields many false corners as described in [16]. In order to overcome the issues in the existing methods the directional filtering approach using Bayer’s pattern is used to improve the quality of the image.

3. EXISTING METHODOLOGY 3.1 Filters

Generally, filters are used to filters the unwanted things or object in a spatial domain or image surface. In digital image processing, mostly the images are been affected by various noises. The main objectives of the filters are applied to improve the quality of the image by enhancing the interoperability of the information present in the image for human visual. A general structure of a filter mask is as follows.

Fig.1 Filtering Mask

Image filtering can be used for many aspects which includes, smoothing, sharpening, noise eliminating and edge detection etc. A filter is defined by a kernel, which represented is a small array and applied to each pixel and its neighbours within an image.

3.2 Gaussian Filter

Gaussian filters are the linear smoothing filter with the weights is selected based on the Gaussian distribution functions. Mainly, these kinds of filters are used to smooth the

-1 -1 -1

-1 N -1

-1 -1 -1



image and to eliminate the Gaussian noises present in the image. This is formulated as follows.

2 2

2 2

1 1( , ) [ ] [ ]2 2 22m nh m n e X

eσ πσ σπσ= .... (1)

From the above equation (1) the Gaussian filter is separable. The Gaussian smoothing filter is very good in noise removal in normal distribution function. This filter is rotationally symmetric the amount of smoothening is all direction. The degree of smoothening is governed by the variance T.

3.3 Vector Median Filter

The Vector Median Filter is similar to the mean filter, which smoothens the data by compute the mean within a windowed subset of the data. Instead of finding the mean for every windowed subset, the VMF finds the median vector. The Vector Median Filter (VMF) and its extensions follow directly from the nonlinear order statistics in that the output of the filter is the lowest ranked vector in the window. The Vector Median Filter orders the color input vectors according to their relative magnitude differences using the Minkowski metric as a distance measure.

arg minm i

k i jmi m kk i ja sS

a a a L= +

= −∈

= −∑ .... (2)

Where given a set of vectors Si = ai−j; ai−j+1,……., ai+j−1; ai+j . j is the window half-width. 3.4 Bilinear Method

3.4.1 Independent interpolation of color planes The simplest method of Demosaicing the images just interpolates each color plane independently using various kinds of interpolation algorithms. One of the common artifacts - a color moiré is present in all demosaicing methods. It results from different space positions of different color sensors. Many demosaicing methods employ the fact that there are twice as many green pixels as red or blue pixels, in order to restore the high-frequency information of the image better. After that, the restored green component is used to interpolate red and blue components. 3.4.2 Color ratios interpolation Interpolation of red and blue colors using the green color is based on some assumptions about correlation of color planes. One of possible assumptions states that ratios of basic color components remain equal within objects of the image. Then, once interpolated the green components, then interpolate ratios of red (or blue) to instead of interpolating red (or blue) colors on their own. Interpolation of green color The missing green pixel is calculated as a linear combination of 4 nearest neighbors of this pixel (the values of these neighbors are known). The weights Ei in the linear combination are calculated from probability that pixel Gi belongs to the same image object as pixel G5.

Green color interpolation

The weights Ei are calculated as follows. Firstly, the concept of directional derivatives for 4 directions (vertical, horizontal, and 2 diagonals) from each point is introduced. Interpolation of red and blue colors using the green color For interpolation of red and blue colors the previously described color ratios interpolation algorithm is used. The ratios are interpolated similarly to green pixels on the previous stage using the weights Åi .

Red color interpolation

4. METHODOLOGY The image processing includes several image-

processing techniques such as filtering, feature extraction noise removal and enhancement of image. Most modern digital photo and video cameras use a mosaic arrangement of photo-sensitive elements. This enables using only one matrix of photo-sensors instead of 3 matrices (one for each basic color component). In such a matrix, the elements sensitive to different basic colors are interleaved. Each element of the matrix stores the information on only one of 3 color components, whereas the output full-color. digital image should contain all 3 basic components (R,G,B) for each pixel. The problem of demosaicing involves the interpolation of color data to produce the full-colored image from the bayer pattern. The demosaicing algorithm interpolates each of color planes at the positions where the corresponding values are missing.

Fig 4.1 Process Flow of Bayer’s Pattern Method

= 2 2 4 4 6 6 8 8

2 4 6 8

E G E G E G E GE E E E+ + ++ + +

- Weight function

R5 =

3 7 911 3 7 9

1 3 7 9

1 3 7 9

R R RRE E E E

G G G GE E E E

+ + +

+ + +

Where Ei - weight function

Input Image

Preprocessed Image

Bayer’s pattern array

Interpolation

R Component G

B R G Component

B Component



The method uses an effective approach by directional interpolation, where the decision of the most suitable direction of interpolation is made on the basis of the reconstructed green component only. Once the choice is made, the red and blue components are interpolated. In this way, the two directional interpolations and the decision concern only one color component and not all the three channels. Moreover, this approach requires the decision only in a half of the pixels of the image, precisely where the sensor did not capture the green samples. Furthermore, since in this case the estimate of the green component after the decision is more accurate, a more efficient reconstruction of red and blue is possible.

A. Directional Green Interpolation Interpolation of Missing G values at B and R Sampling Positions The first step is to reconstruct the green image along horizontal and vertical directions. A five-coefficient FIR filter is to interpolate the Bayer samples; the green signal is sub sampled with a factor of 2. In the frequency domain,

( ) ( ) ( )1 12 2SG w G w G w π= + − .... (3)

Where G(ω) and Gs(ω) denote the Fourier transform of the original green signal and of the down-sampled signal, respectively. Therefore, if G(ω) is band-limited to | ω| < П/2, the ideal interpolation filter to perform the reconstruction would be

( ) ( )2 *idwH w re c t π= .... (4)

Since, it eliminates the aliasing component 1/2 G (ω- П). The only FIR filter with three coefficients that can apply to Gs(ω) without modifying the average value of the samples is h0 = [0.5 1 0.5].

( ) ( ) ( )0*SG w w wHG∧

= After filtering

( ) ( ) ( ) ( )0 0

1 12 2

w w w wG GH Hπ= + − .... (5)

Where the second term denotes the aliasing component.

In a green-red row, the red component is sampled with an offset of one sample with respect to the green signal. Therefore, its Fourier transforms results

( ) ( ) ( )1 12 2SR w R w R w π= + − .... (6)

Where R(ω) is the fourier transform of the original red signal. If interpolate it with a filter and then add the resulting signal to (5)

( ) ( ) ( ) ( ) ( )

( ) ( ) ( ) ( )

0 0

1 1

1 12 2

1 12 2

G w G w w G w w

R w w R w w

H H

H H

π

π

∧

= + −

+ − −

.... (7)

Reminding us that R(ω)-G(ω) is slowly varying, if h1 is designed such that at low frequencies and H1(ω) ≅ H0(ω) at high frequencies, ( ) ( )1R w H w ≅ ( ) ( )1G w H w .... (8)

( ) ( )0G w H wπ− ≅ ( ) ( )1R w H wπ− .... (9)and (7) could be approximated as

( )G w∧

≅ ( ) ( ) ( ) ( )0 1

1 12 2

G w w R w wH H+ .... (10)

A good choice for a filter h1 that respects the constraints

(8) and (9) is the five-coefficient FIR [-0.25 0 0.5 0 -0.25]. The missing green sample G0 is estimated as

( )0 2 0 20 1 1

12 2 2

R R R RG R G G∧

−−

− −= + − + − ... (11)

That is, this reconstruction can also be considered as a bilinear interpolation of the R – G difference, where the unknown values R1 and R-1 are estimated as (R0 + R2) / 2 and (R0+R-2) / 2, respectively. The interpolation of the green values in the blue-green rows and the interpolation along the columns follow the same approach. Once the green component has been interpolated along both horizontal and vertical directions and two green images have been produced, a decision has to be made to select the filtering direction that gives the best performance. Let GH and GV be the two interpolated green images. For each image, in every red or blue location, the chrominance values (or ) in a red pixel and (or) in a blue pixel are calculated.

⎧⎪⎨⎪⎩

⎧⎪⎨⎪⎩

−=

−

−=

−

, ,

, ,

, ,

, ,

( , )( , )

( , )

( , )( , )

( , )

Hi j i j

H Hi j i j

Vi j i j

V Vi j i j

R G if i j is a r ed lo c a t ionC i j

B G if i j is a b lu e lo c a t ion

R G if i j is a r ed lo c a tionC i j

B G if i j is a b lu e lo c a tion

( ) ( )( ) ( )

, , ( , 2)

, , ( 2, )H H H

v v v

D i j C i j C i j

D i j C i j C i j

= − +

= − +

Where i and j indicate the row and the column of the pixel ( i , j), 1 ≤ i ≤ M, 1 ≤ j ≤ N ( M and N denote the height and the width of the image, respectively). Note that CH and CV are not defined in the green pixels. Next calculate the gradients of the chrominance and, precisely, the horizontal gradient for CH and the vertical one for CV ( , ) ( , )H Vi j and i jδ δ and as the sum of the gradients DH and DV belonging to a sufficiently large neighborhood of (I, j).With a square window, both the classifiers are computed considering the same number of gradients based on the red chrominance and the same number of gradients based on the blue chrominance. The two classifiers ( , ) ( , )H Vi j and i jδ δ give an estimate of the local variation of the color differences along the horizontal and vertical directions, respectively, and they



can be used to estimate the direction of the edges. For example, if the value of is lower than, it is likely that there is a horizontal edge instead of a vertical one. For all the red and blue pixels, then estimate the green values using the following criterion: If ( , ) ( , )V Hi j i jδ δ< then , ,

ˆ Vi j i jG G=

Else , ,ˆ Hi j i jG G=

Considering the known green samples, a full resolution green image is estimated. An additional improvement can be included in this procedure. Usually, in natural images, the majority of the edges and the details present cardinal orientations. Therefore, if the pixel ( i , j ) is placed in a detailed area, during the estimation of the green values ,

ˆi jG , it can result

preferable to give more weight to the gradients and of the pixels in the same row and column of (i , j). This can be accomplished by weighting these gradients two or three times more than the other gradients when calculate ( , ) ( , )H Vi j and i jδ δ . In our implementation a weight of 3 is used to this purpose. Interpolation of Missing R/B Values at G Sampling Positions

To interpolate the missing R and B values at a G sampling position. For concreteness and without loss of generality, let us examine a subcase. Subcase 1: G sampling position with horizontal R and vertical B neighbors, as illustrated below. Note that, by now all four neighboring green values ( , ),( , ),( , ),( , )g g gc gc gn gn gs gsh v h v h v h vω ω interpolated in both directions are available. Using these reconstructed green values and the original sample values compute,

( )

( )

( )

( )

12121212

w cr c c g w g c

w cr c c g w g c

n sb c c g n g s

n sb c c g n g s

h G h hR R

v G v vR R

h G h hB B

v G v vB B

= + − + −

= + − + −

= + − + −

= + − + −

Since blue color is not sampled at all in the current row and red samples are completely missing in the current column, maintaining primary consistency is difficult when estimating

rcv and bch . The vertical R interpolation rcv has to use red samples of the horizontal neighbors

,w cR R and (1/2)( )rc c w gw c gcv G R v R v= + − + − , which is in conflict with the underlying assumption of vertical structure. The best one can do here is to fully utilize available vertical information of the neighboring columns.

The green estimates gv ω and gcv associated with

w cR and R and are used to estimate rcv . It is important to

realize that gv ω and gcv is estimated under the hypothesis of vertical structure. The influence of the vertical structure to the missing red value in the current column is factored in by assuming that the difference image between the red and green channels is reasonably smooth in the small locality. Subcase 2: Consider the following mosaic configuration of G sampling position with horizontal B and vertical R neighbors: The estimation of missing R and B for G sampling position

with horizontal B and vertical R neighbors can be derived symmetrically to Subcase 1 under the same rationale.

( )

( )

( )

( )

12121212

w cb c c g w g c

w cb c c g w g c

n sr c c g n g s

n sr c c g n g s

h G h hB B

v G v vB R

h G h hR R

v G h vR R

= + − + −

= + − + −

= + − + −

= + − + −

One of the common artifacts a color moiré present in demosaicing color mosaic image using directional filtering approach. It results from different space positions of different color sensors. Many demosaicing methods employ the fact that there are twice as many green pixels as red or blue pixels, in order to restore the high-frequency information of the image better. After that, the restored green component is used to interpolate red and blue components. 5. SIMILARITY MEASURES 5.1 Mean Squared Error (MSE) Mean square error is given by

[ ]2

1 1

1 ( , ) ( , )M N

i iMSE g i j f i j

MN = == −∑ ∑ .... (12)

Where M and N are the total number of pixels in the horizontal and the vertical dimensions of image, g denotes the Noise image and f denotes the filtered image. 5.2 Peak Signal to Noise Ratio (PSNR) The peak Signal to Noise ratio is calculated by:

2

1 02 5 51 0 lo gP S N RM S E

⎛ ⎞= ⎜ ⎟

⎝ ⎠ .... (13)

For the image quality measures, if the value of the PSNR is very high for an image of a particular noise type then is best quality image.

ω ω

, ,, , , ,

, ,

n g n g n

g g w c c g c g c

s g s g s

B h vR h v G R h v

B h v



6. ALGORITHM Input : Input image from IDB Output : Pre-processed image

7. EXPERIMENTATION AND RESULTS The proposed model is experimented with a set of ten test of face images of size (256 x 256) or (512 x 512) color image is has process with the bayer pattern array. The Bayer pattern filters the image into Bayer pattern image. In this image each and every pixel is associated with either Green, Blue or Red colors. The output image is as shown in the following figure,

Fig 3. (a) Original Image (b) Bayer pattern image

The chrominance values R - GH (or R – GV) in a red pixel, and B - GH (or B - GV ) in a blue pixel are calculated for the horizontal and red regenerated green images. Also calculate color gradient value of the chrominance DH and DV. For each red or blue pixel, define the classifiers ( , )H i jδ and ( , )V i jδ as the sum of the gradients DH and DV belonging to a sufficiently large neighborhood of (i, j). Estimate the local variation of the color differences along the horizontal and vertical directions, respectively and they can be used to estimate the direction of the edges. If the value of Hδ is lower than Vδ , it is likely that there is a horizontal edge instead of a vertical one. The experimented results of the proposed method and the existing method with the

computation of MSE and PSNR values are presented in the following tables.

Table 1. MSE Comparison

Image No

Bilinear Method (BM)

Gaussian Filter (GF)

Vector Median Filter

(VMF)

Bayer’s Pattern Method (BPM)

1 55.59 85.47 3.21 0.92

2 41.36 73.59 3.05 1.01

3 39.64 78.38 2.82 0.94

4 39.81 81.60 4.64 2.24

5 47.83 98.31 6.40 1.65

6 46.13 79.33 4.79 1.18

7 46.89 85.62 5.36 1.11

8 45.95 84.01 5.44 1.22

9 47.55 89.62 5.78 1.16

10 49.05 105.43 10.50 2.85 From the table 1 shows the experimented values

obtained from different preprocessing methods. It shows the selected face image from the database. The performance was evaluated using the Mean Square Error (MSE) and Peak Signal Noise Ratio (PSNR) in order evaluates the quality of the image. By the analysis of the values in the table the Bayer’s Pattern Method is better with less MSE and high PSNR values. In order to evaluate the performance of the Bayer’s Pattern Method considered the obtained results with the existing bilinear method, Gaussian Filter, vector median filter are shown in the following table 2.

Table 2. PSNR Comparison

Image No

Bilinear Method

(BM)

Gaussian Filter (GF)

Vector Median Filter

(VMF)

Bayer’s Pattern Method (BPM)

1 12.6046 40.327 54.577 59.9996

2 13.4795 40.977 54.799 59.5843

3 13.6636 40.703 55.144 59.7602

4 13.6422 40.528 52.979 56.1509

5 12.8487 39.720 51.586 57.4683

6 13.0053 40.651 52.847 58.9402

7 12.9346 40.320 52.358 59.173

8 13.0226 40.402 52.287 58.7802

9 12.8742 40.122 52.026 58.9878

10 12.7391 39.416 49.433 55.1045

From the below figure 4 shows the pictorial representation of the performance evaluated. By analysing the obtained results the proposed model produced the best results. Hence the Bayer’s Pattern method is an efficient one.

Step 1: Select input image of size 256 x 256 from the image database

Step 2: Convert into Bayer pattern images using Bayer pattern array. Step3: Estimate the difficult missing green both horizontal GH and GV. Step 4: Calculate the chrominance values R – GH in red, and B - GH in blue pixel. Step5: Calculate color gradient value of the chrominance DH and DV. Step6: For each red or blue pixel, define the classifiers ( , )H i jδ and ( , )V i jδ Step7: Estimate the local variation of the color differences along the horizontal and vertical directions, respectively. If the value of Hδ is lower than Vδ , it is likely that there is a horizontal edge instead of a vertical one. Step8: Repeat the above steps to estimate a full resolution

green image G . Step9: Then green channel has been reconstructed, interpolate the red and blue components. Step10: Finally, the original color image has been outputted.



Fig 4. Performance Evaluation

8. CONCLUSION In this paper, the demosaicing of images based on directional filtering and a posteriori decision has been presented. The experimental result proves the effectiveness of this approach, providing good PSNR values when compared to existing methods. The performances of PSNR values of proposed Bayer’s Pattern method when compared to existing methods Bilinear Method, Gaussian Filter and Vector Median Filter are investigated independently. The proposed Bayer’s Pattern method produces better results with 58.39% accuracy compared with existing methods gives 13.08% accuracy for Bilinear Method, Gaussian Filter with 40.31% accuracy and Vector Median Filter with 52.80% accuracy. Moreover, the computational cost of the algorithm is very low. Therefore, the proposed algorithm candidates itself for implementation is in simple low-cost cameras or in video capture devices with high values of resolution and frame rate. The proposed scheme is capable of achieving at least comparable and often better performance than existing iterative demosaicing techniques. 9. REFERENCES [1] Seyed Pooya Ehsani, Hojjat Seyed Mousavi,

Babak.H.Khalaj “Chromosome Image Contrast Enhancement Using Adaptive, Iterative Histogram Matching” 978-1-4577-1535- 8/11/ 2011 IEEE.

[2] Min Liu, Peizhong Liu “Image Enhancement Algorithm for Video Based On Multi-Dimensional Biomimetic Informatics” 978-0-7695-4647-6/12 2012 IEEE DOI 10.1109/ICCSEE.2012.244

[3] R.K. Jha, P.K. Biswas, B.N. Chatterji “Contrast enhancement of dark images using stochastic resonance” IET Image Processing, 2012, Vol. 6, Iss. 3, pp. 230–237; doi: 10.1049/iet-ipr.2010.0392.

[4] Sudharsan Parthasarathy, Praveen Sankaran “Fusion Based Multi Scale Retinex with Color Restoration for Image Enhancement” IEEE,978-1-4577-1583-9/ 12/

2012. [5] Deepak Ghimire and Joonwhoan Lee “Nonlinear

Transfer Function-Based Local Approach for Color Image Enhancement” IEEE Transactions on Consumer Electronics, Vol. 57, No. 2, May 2011.

[6] Hong ZHANG，Qian ZHAO，Lu LI, Yue-cheng LI, Yuhu YOU “Multi-scale Image Enhancement Based

on Properties of Human Visual System” 978-1-4244-9306-7/11/ 2011 IEEE.

[7] Adin Ramirez Rivera, Byungyong Ryu, and Oksam Chae “Content-Aware Dark Image Enhancement through Channel Division” IEEE Transactions on image processing, VOL. 21, NO. 9, September 2012.

[8] Zhang Chaofu, MA Li-ni, Jing Lu-na “Mixed Frequency domain and spatial of enhancement algorithm for infrared image” 978-1-4673-0024-7/10/2012 IEEE.

[9] B. E. Bayer, “Color Imaging Array,” U.S. Patent 3 971 065, 1976.

[10] W. Lu and Y.-P. Tan, “Color filter array demosaicing: New method and performance measures,” IEEE Trans. Image Processing, vol. 12, no. 10, pp. 1194–1210, Oct. 2003.

[11] C. A. Laroche and M. A. Prescott, “Apparatus and Method for Adaptively Interpolating a Full Color Image Utilizing Chrominance Gradients,” U.S. Patent 5 373 322, 1994.

[12] R. H. Hibbard, “Apparatus and Method for Adaptively Interpolating a Full Color Image Utilizing Chrominance Gradients,” U.S. Patent 5 382 976, 1995.

[13] K. Hirakawa and T. W. Parks, “Adaptive homogeneity-directed demosaicing algorithm,” IEEE Trans. Image Processing, vol. 14, no. 3, pp. 360–369, Mar. 2005.

[14] B. K. Gunturk, Y. Altunbasak, and R. M. Mersereau, “Color plane interpolation using alternating projections,” IEEE Trans. Image Processing, vol. 11, no. 9, pp. 997–1013, Sep. 2002.

[15] D. H. Brainard,“Bayesian method for reconstructing color images from trichromatic samples,” in Proc. IS&T 47th Annu. Meeting, 1994, pp. 375–380.

[16] H. J. Trussell and R. E. Hartwig, “Mathematics for demosaicking,” IEEE Trans. Image Processing, vol. 3, no. 4, pp.485–492, Apr. 2002.

AUTHORS PROFILE

Dr. S. Pannirselvam was born on June 23rd 1961. He is working as Associate Professor and Head, Department of Computer Science in Erode Arts & Science College (Autonomous), Erode, Tamilnadu, India. He is research supervisor M.Phil and Ph.D programmes. His area of interests includes, Image Processing, Artificial Intelligence, Data Mining, Networks. He has presented more than 15 papers in National and International level conferences. He has published more than 18 papers in International journals.

S.Prasath currently pursuing Ph.D as a full time research scholar under the guidance of Dr.S.Pannirselvam at Department of Computer Science, Erode Arts & Science College (Autonomous), Erode, Tamilnadu, India. He has obtained his Masters degree in Software Engineering from M.Kumarasamy college of Engineering, Karur under Anna University, Chennai and M.Phil degree in Computer Science. His area of interests includes, Image Processing and Data Mining.He has presented 2 papers in National and 1 International level conference.




An Agent-Based Framework for Virtual Machine

Migration in Cloud Computing

Somayeh Soltan Baghshahi *

Computer Engineering Department

Islamic Azad University, North Tehran Branch

Tehran, Iran

Sam Jabbehdari



Tehran, Iran

Sahar Adabi



Abstract— Cloud computing is a model for large-scale distributed

computing, which services to customers be done through a

dynamic virtual resources with high computational power of

using the Internet. The cloud service providers use different

methods to manage virtual resources, that to use of autonomous

nature of the intelligent agents, it can improve quality of service

in a cloud distributed environment.

In this paper, we design a framework by using of the multiple

intelligent agents, which these agent interactions with together

and they manage to provide the service. Also, In this framework,

an agent is designed to improve the migration technique of

virtual machines.

Keywords- Cloud Computing; Virtualizaion; Virtual Machine

Migration; Agent-Based Framework

I. INTRODUCTION

Cloud computing is an emerging new paradigm for hosting

services over the internet. Cloud computing offers

infrastructure as a service, platform as a service and storage as

a service to the cloud users. Cloud users are charged based on

their service usage. The cloud computing services are available

at anywhere, anytime, only we have to need internet

connectivity. To improve the utilization of cloud resources we

use virtual machines. The virtual machine is a software

implementation of a computing environment in which

operating system or program can be installed and run [1].

Cloud computing [2] has currently attracted considerable

attention from both the industrial community and the academic

community. Virtualization provides an abstraction of hardware

resources enabling multiple instantiations of operating systems

to run simultaneously on a single physical machine. Another prominent advantage of the Virtualization is the

live migration technique [3,5] which refers to the act of migrating a virtual machine from one physical machine to another even as the virtual machine continues to execute. Currently, live migration has become a key ingredient behind the management activities of cloud computing system to achieve the goals of load balancing, energy saving, failure recovery, and system maintenance [4].

II. RELATED WORK

Virtualization is the key technology that enables the emerging

cloud computing paradigm [8][9][10], because it allows

resources to be allocated to different applications on demand

and hides the complexity of resource sharing from cloud users.

VMs are generally employed in different types of cloud

systems as containers for hosting application execution

environments and provisioning resources. For example, in

Infrastructure-as-a-Service (IaaS) clouds [3], VMs are directly

exposed to users to deliver a full computer infrastructure over

the internet; In Platform-as-a-Service (PaaS) clouds [13], VMs

are also used by the clouds internally to manage resources

across the application execution platforms delivered to users.

VM migration is a unique capability of system Virtualization

which allows an application to be transparently moved from

one physical host to another and to continue its execution after

migration without any loss of progress. It is generally done by

transferring the application along with its VM’s entire system

* Corresponding author






state, including the state in CPU, memory, and sometimes disk

too, from the source host to the destination host. VM migration

is an important means for managing applications and resources

in large scale virtualized data centers and cloud systems. It

enables resource usage to be dynamically balanced in the entire

virtualized system across physical host boundaries, and it also

allows applications to be dynamically relocated to hosts that

can provide faster or more reliable executions [11].

Chih et al [16] have proposed an agent-based service migration

framework in the cloud computing environment to manage

resources and monitor the behavior of the system.

III. VM MIGRATION

Virtual machine migration takes a running virtual machine and

moves it from one physical machine to another. This migration

must be transparent to the guest operating system, applications

running on the operating system, and remote clients of the

virtual machine. It should appear to all parties involved that the

virtual machine did not change its location. The only perceived

change should be a brief slowdown during the migration and a

possible improvement in performance after the migration

because the VM was moved to a machine with more available

resources [6].

Live VM migration technologies have proven to be a very

effective tool to enable data center management in a

nondisruptive manner. Both Xen and VMware adopts pre-

copying algorithm for VM live migration in a memory-to-

memory approach [3]. In the approach, physical memory image

is pushed across network to the new destination while the

source VM continues running.

Pages dirtied during the migration must be iteratively

re-sent to ensure memory consistency. By iterative it means

that pre-copying occurs in several rounds and the data to be

transmitted during a round are the dirty pages generated in the

previous round. The pre-copying phase terminates (1) if the

memory dirtying rate exceeds the memory transmission rate; or

(2) if the remaining dirty memory becomes smaller than a pre-

defined threshold value; or (3) if the number of iterations

exceed a given value; or (4) the network traffic exceeds a

multiple of the VM memory size. After several rounds of

synchronization, a very short stop-and-copy phase is performed

to transmit the remaining dirty pages. As the data transferred is

relatively small, this mechanism results in a nearly negligible

best-case migration downtime.

We note that the performance of live VM migration is affected

by many factors. First of all, the size of VM memory has a

main effect on the total migration time and network traffic.

Secondly, the memory dirtying rate, which reflects the memory

access pattern of different applications, impacts the number of

iteration rounds and data transferred in each

pre-copying round, and hence indirectly affects the migration

latency and network traffic. Thirdly, the network transmission

rate together with the configuration of migration algorithm is

also crucial to migration performance [7].

IV. AGENT-BASED SYSTEMS[14]

Agent-based systems are software systems that use agents to

perform problem solving or other computational tasks. In

agent-based systems, the system’s task is assigned to

autonomous software entities called agents which in turn

cooperate with each other in order to complete it.

Agents are a special category of computer programs that in

contrast to conventional programs have the ability to act

autonomously. An agent is programmed as to be able to

perceive a specific environment and additionally acts upon it.

Moreover, each agent has a specific objective, a goal, and thus

it must take action upon its environment in order to achieve it.

An agent’s autonomous behavior derives from the ability of

being both proactive and reactive. By being proactive, an

agent adjusts its behavior and plans its actions as to achieve its

initial goal, while by being reactive it is able to respond to

changes in its environment in a timely manner. It is also

possible for agents to learn by gathering information from

their environment and their previous actions.

This information is then stored internally in the agent in the

form of beliefs which can ultimately affect the agent’s

behavior.

In addition to being autonomous, agents have the ability to

engage in social interactions with other agents. Their social

abilities enable them to exchange information, to cooperate in

order to achieve their goals and to coordinate their actions.

Agents can also have different roles in a system and may

influence the behavior of other agents or even control them by

requesting specific actions.

Their unique design and its provided features along with their

social abilities make agents suitable for creating complex

systems [15]. Using agents can simplify the design and

implementation of such systems, as not all possible links,

interactions and states will have to be considered. Instead,

agents can be programmed with specific behaviors that will

enable them to deal with unknown states and interactions as

they occur. Furthermore, agents can either be used

individually, by assigning each one of them to work on a

specific aspect of the problem or together, by letting them

cooperate as to solve a problem in a distributed fashion.

A. Agent-Based Computing[13]

An agent is a computer system that is capable of autonomous

(independent) actions, that is, deciding for itself and figuring

out what needs to be done to satisfy its design objectives [12].

A multi-agent system consists of a number of agents, which

interact with one another [12]. To successfully interact, agents

require the ability to cooperate, coordinate, and negotiate with

each other. Cooperation is the process when several agents

work together and draw on the broad collection of their

knowledge and capabilities to achieve a common goal.

Coordination is the process of achieving the state in which

actions of agents fit in well with each other. Negotiation is a

process by which a group of agents communicate with one

another to try to come to a mutually acceptable agreement on

any matter.




B. Agent-Based Cloud Computing

Some of the essential characteristics of cloud computing

include resource pooling and resource sharing. In the clouds,

computing resources are pooled to serve multiple consumers,

and applications and data are available to and shared by a broad

group of cross-enterprise and cross-platform users. Resource

pooling and sharing involve 1) combining resources through

cooperation among cloud providers, 2) mapping, scheduling,

and coordination of shared resources, and 3) establishment of

contracts between providers and consumers. In agent-based

cloud computing, cooperation, negotiation, and coordination

protocols of agents are adopted to automate the activities of

resource pooling and sharing in clouds.

V. PROPOSED FRAMEWORK

In this paper, we design a framework that consists of several

clouds. In this framework uses of intelligent agents that each of

these agents is independent and they are able to communicate

with each other. This framework is done in two phases: design

and service.

1) Design Phase

In this phase is determined the number of physical machines

and virtual machines, service type and arrangement of services

on virtual machines and virtual machine clustering. The

purpose of this phase is clustering and resource allocation for

virtual machines that lead to reduced network traffic, efficient

use of the bandwidth and increase the performance. In this way,

the number of services in the cloud is considered constant.

2) Service Phase

After the design phase and preparation of the system, this phase

begins. According to customers' requests, these requests are

responded locally and remotely. If the number of customer

requests is high and requires virtual machines are migrated, the

migration done in this phase.

A. The Components of the Proposed Framework

In this approach, a prototype has been developed which

consists of four clouds and each cloud contains multiple

physical machines, an application, a network storage space for

storing data and a set of agents.

The application consists of two main parts:

1) Cloud manager:

a) Storing information relating to the physical machines

and virtual machines in the database

b) Service Management in the Cloud

2) Agent manager:

a) Agents distribution of physical machines

b) Agents management and communicate with each

other

Fig.1. Proposed Framework




In this way, all the clouds are directly connected with each

other by VPN1 protocols.

Applications in the cloud are associated with other clouds.

The proposed framework is shown in Figure 1.

B. The Agents in theFramework

In this prototype, five agents work together:

1) Clustering Agent (CA) This agent performs virtual machines clustering, so that,

initially, the service will be given to this agent. The task of this

agent is division of service to the different sections and each

section is placed on a virtual machine. After the service

division and determined features, such as the size of sub-

sections, the virtual machines are created in accordance with

this sub-section. Each sub-section has a priority which this

priority will be determined with an integer, then the service

components are placed according to their priorities in a

priority queue. One of the virtual machines is randomly

selected as the root that is responsible for the coordination

between the components. After selecting the root, different

parts of a service are removed from the queue in FIFO2 mode

and they are placed on virtual machines. This is shown in

Figure2.

Fig.2. Virtual machines clustering

Information related to clusters and virtual machines in a

service is stored within the local database in the cloud storage.

If in the service phase, several clusters (Service) were

prepared simultaneously for migration, first, then based on the

number of requests related to each service, they are placed in a

priority queue and they are removed sequentially from the

queue.

2) Request Management Agent (RMA) This agent is responsible for requests analysis that consists of

two parts as follows:

a) Request Monitoring (RM)

This section all the information about each request will be

saved in a database. Characteristics of these requests, which

must be stored is as follows:

- Requests Identification (type of service) (IDs)

- Requests Source (RS)

- Requests start time (RT)

- Requests Position (Remote / Local) (RP)

1 Virtual Private Network 2 First In First Out

When the number of remote requests table is larger than the

threshold ( ), a message is sent to the RA section, that

requests must be analyzed.

The clouds are able to respond to remote requests. If the

number of requests has exceeded than the threshold, for

establishing the quality of the service must be analyzed

requests and new destination to be determined and then a

series of virtual machines will be migrate the new destination.

b) Request Analyzer (RA) In this section, after receiving the message of RA, the table of

requests is analyzed, then service with highest remote request

will be determined.

For example, Figure3 shows the analysis of the requests.

Fig.3. Example of requests analyzing

When service with a highest remote request is determined, this

agent communicates with the PM agent and sends the required

information and the PM agent determines the new location of

service.

3) Monitoring Agent (MoA)

This agent is responsible for monitor the entire system, that

consists of two main parts:

a) Vm Snapshot

In this section, in the time intervals of each virtual machine are

taken a SnapShot and the vm snapshot is stored in the

database. The time intervals are determined by Predictor

agent.

b) Log

In this section, the entire system logs are stored in the

database. During of the migration, If the system was damaged

the Log will be used. Also, this section is used to predict the

time Snapshot.

4) Prediction Agent (PA) This agent is responsible for Analysis for Log Table. By

applying this table, the migration time is obtained, then the

agent can predict a perfect time for the VM Snapshot.

5) Migration Agent (MA) This agent is responsible for migration of virtual machines

between physical machines that consists of two main parts:

a) Pre-Migration(PM)

This section is responsible for obtaining the necessary

information about service, that is ready for migration. This

information can be obtained by communicate with the source

application.




Then, according to the necessary information, it uses of the

Analytical Hierarchy Process (AHP) for the find of

destination.

b) Migration

After receiving the message of the Pre-Migration section, this

section will perform the migration. In proposed method, we

use the greedy algorithm for virtual machines migration. In

this method, the selection procedure for virtual machines

migration is memory size of the virtual machines and storage

space of the destination physical machine. So that, at each

step, a comparison is done between the memory size of virtual

machine and storage space of the physical machine, If the

memory size of virtual machine was smaller than storage

space of the physical machine, it will be selected for

migration.

VI. CONCLUSION

According to studies, currently, there are many challenges and

failings in providing service to customers through the internet,

Such as the lack of quality of service, lack sufficient

computing resources, geographical distance between clients

and service providers, lack enough bandwidth and high

volume of data transferred. Cloud computing using

virtualization technology is improved these weaknesses. In the

proposed framework, a discipline is organized through the

intelligent agents between the server components, that the

service will be provided using these agents. The migration

agent performs migration by two methods of hierarchical and

greedy algorithm. In the hierarchical approach is determine the

most appropriate location for the destination and in the greedy

method is selected a virtual machine for migrate.

Virtual machine migration techniques to increase the

flexibility and scalability of data center in the cloud

environments. Also, the use of multi-agent environment makes

the problem of complexity is solved in distributed

environment.

REFERENCES

[1] Bhaskar Prasad Rimal, Eunni Choi, Ian Lumb “a taxonomy and Survey of Cloud Computing Systems” 2Fifth International Joint Conference on INC, IMS and IDC 2009.

[2] M. Armbrust, A. Fox, R. Griffith, A. Joseph, R. Katz, A. Konwinski, G. Lee, D. Patterson, A. Rabkin, I. Stoica, et al., “A view of cloud computing,” Communications of the ACM, vol. 53, no. 4, pp. 50–58, 2010.

[3] M. Nelson, B. Lim, and G. Hutchins, “Fast transparent migration for virtual machines,” in Proceedings of the annual conference on USENIX Annual Technical Conference, p. 25, 2005.

[4] K. Ye, D. Huang, X. Jiang, H. Chen, and S. Wu, “Virtual machine based energy-efficient data center architecture for Cloud computing: a performance perspective,” in Proceedings of the 2010 IEEE/ACM International Conference on GreenComputing and Communications (GreenCom), pp. 171– 178, 2010.

C. Clark, K. Fraser, S. Hand, J. G. Hansen, E. Jul, C. Limpach, I. Pratt, and A. Warfield, “Live migration of virtual machines,” in Proc. USENIX Symposium on Networked Systems Design and Implementation (NSDI’05), Berkeley, CA, USA, pp. 273–286, 2005.

[5] M.Nelson, B.Lim, “Fast Transparent Migration for Virtual Machines,”USENIX Annual Technical Conference,April – 2005.

[6] H.Lio, H.Jin, h.Xu, X.Liao, “Performance and energy modeling for live migration of virtual machines ,” Cluster Computing,December- 2011.

[7] Amazon Elastic Compute Cloud (Amazon EC2), URL: http://aws.amazon.com/ec2/.

[8] Windows Azure Platform, URL:

http://www.microsoft.com/windowsazure/.

[9] Google App Engine, URL:http://code.google.com/appengine/.

[10] Y.WU, M.Zaho,“ Performance Modeling of Virtual Machine Live Migration,”IEEE International Conference on Cloud Computing (CLOUD), 2011.

[11] M. Wooldridge, An Introduction to Multiagent Systems, second Ed. John Wiley & Sons, 2009.

[12] K. Mong Sim,”Agent-Based Cloud Computing”,IEEE Transactions on Services Computing, Vol. 5, No. 4, October-December 2012.

[13] A. Vichos,”Agent-based management of Virtual Machines for Cloud infrastructure”, Master of science, Computer Science, School of Informatics University of Edinburgh,2011.

[14] N. R. Jennings,” On agent-based software engineering”, Artificial Intelligence, 117(2):277–296, 2000.

[15] Ch. Fan, W.Ang, Y.Chang, “Agent-based Service Migration Framework in Hybrid Cloud”, International Conference on High Performance Computing and Communications, IEEE, 2011.




Migration of Virtual Clusters with Using Weighted

Fair Queuing Method in Cloud Computing

Leila Soltan Baghshahi


Islamic Azad University, South Tehran Branch

Tehran, Iran

Ahmad Khademzadeh

Education and National International Scientific Cooperation

Department

Iran Telecommunication Research Center (ITRC)

Tehran, Iran

Sam Jabbehdari



Tehran, Iran

Abstract— Load Balancing, Failure Recovery and Quality of

Services, portability are some of the advantages in virtualization

technology and cloud computing environment.

In this environment, with uses the feature of Encapsulation,

virtual machines together is considered as a cluster, that these

clusters are able to provide the service in cloud environments.

In this paper, multiple virtual machines are considered as a

cluster. These clusters are migrated from a data center to another

data center with using weighted fair queuing. This method is

simulated in CloudSim tools in Eclipse and Java programming

language. Simulation results show that the bandwidth parameter

plays an important role for the virtual machine migration.

Keywords-Cloud Computing; Virtualizaion; Virtual Cluster;

Live Migration

I. INTRODUCTION

Virtual machine (VM) technology has recently emerged as an

essential building block for data centers and cluster systems,

mainly due to its capabilities of isolating, consolidating and

migrating workload [1]. Altogether, these features allow a data

center to serve multiple users in a secure, flexible and efficient

way. Consequently, these virtualized infrastructures are

considering a key component to drive the emerging Cloud

Computing paradigm [2].

Migration of virtual machines seeks to improve manageability,

performance and fault tolerance of systems.

Cloud computing [3] has currently attracted considerable

attention from both the industrial community and academic

community. In this new computing paradigm, all the resources

are delivered as the services (Infrastructure Service, Platform

Service, and Software Service) to the end users via the

Internet. Virtualization [1, 4] is a core technique to implement

the cloud computing paradigm. Virtualization provides an

abstraction of hardware resources enabling multiple

instantiations of operating systems to run simultaneously on a

single physical machine. Another prominent advantage of the

virtualization is the live migration technique [4, 6] which

refers to the act of migration a virtual machine from one

physical machine to another even as the virtual machine

continues to execute. Currently, live migration has become a

key ingredient behind the management activities of cloud

computing system to achieve the goals of load balancing,

energy saving, failure recovery, and system maintenance [7].

Virtual Cluster (VC) [8, 9] is a group of virtual machines

configured for a common purpose.

System virtualization is a powerful platform for provisioning

applications and resources in the emerging computer systems

such as utility data centers and cloud systems. Live VM

migration is an important tool for managing such systems in

various critical aspects such as performance and reliability.

Understanding the role that the resource availability plays on

the performance of live migration can help us make better




decisions on when to migrate a VM and how to allocate the

necessary resources.[10]

II. VIRTUAL MACHINE

One computer containing multiple operating systems loaded

on a single PC, each of which functions as a separate OS on a

separate physical machine [4].

A virtual machine [5] behaves exactly like a physical

computer and contains its virtual (i.e. Software-based) CPU,

RAM hard disk and network interface card (NIC).

A. Benefits of Virtual Machine[13]

1) Isolation:

While virtual machines can share the physical resources of a

single computer, they remain completely isolated from each

other as if they were separate physical machines.

If, for example, there are four virtual machines on a single

physical server and one of the virtual machine crashes, the

other three virtual machines remain available. Isolation is an

important reason why the availability and security of

applications running in a virtual environment is far.

2) Encapsulation:

A virtual machine is essentially a software container that

bundles or encapsulates a complete set of virtual hardware

resources, as well as an operating system and all its

applications, inside a software package. Encapsulation makes

virtual machines incredibly portable and easy to manage.

3) Hardware Independence:

Virtual machines are completely independent from their

underlying physical hardware. For example, you can configure

a virtual machine with virtual components (e.g. CPU, network

card, SCSI controller) that are completely different from the

physical components that are present on the underlying

hardware. Virtual machines on the same physical server can

even run different kinds of operating systems (Windows,

Linux, etc.).

III. VIRTUALIZATION IN THE CLOUD

Virtualization is a means to create one or more instances of

a real model, such that users are not aware of its virtual. In the

computer world, creating a model (logical layer) of the system

hardware and run programs on a virtual model is known as

virtualization. In other words, has created a model similar to

the model actual and distinct from other parts of the system.

This virtual model is called a virtual machine [1].

Virtualization technology has many advantages such as

security isolation, hiding heterogeneous hardware reliability

and so on. This technology provides greater efficiency of

computing resources, improved scalability, reliability and

availability [5].

Virtualization technology forms the core of the cloud

computing model. With this technology, the physical

machines are converted into multiple Virtual Machine that

each of them will be responsive to the needs of multiple

customers.

The virtual machines have become to a common level of

abstraction and a unit for providing applications, because they

are the least common element between customers and service

providers.

The use of virtual machines is not sufficient to meet the needs

of customers within the data centers; you can't remain only

with these tools in the competitive market . Theref

-

. This technique is c

purposes such as load balancing, repair of servers, failure

recovery, and increase availability and so on.

VM migration techniques include techniques Pre-Copy, the

Post-Copy, Three-Phase, CR / TR, Heterogeneous, and are

aware of the dependency. Evaluation parameters of migration

techniques include:

Total Migration Time

Downtime

The volume of transmitting data

Overhead

A. VM Live Migration

Live migration is a technology with which an entire running

VM is moved from one physical machine to another.

Migration at the level of an entire VM means that active

memory and execution state are transferred from the source to

the destination. This allows seamless movement of online

services without requiring clients to reconnect [11].

Live VM migration technologies have proven to be a very

effective tool to enable data center management in a non-

disruptive manner. Both "Xen" and VMware adopts pre-

copying algorithm for VM live migration in a to memory

approach [16, 17], as shown in Fig. 1. In the approach,

physical memory image is pushed across network to the new

destination while the source VM continues running.

Pages dirtied during the migration must be iteratively re-sent

to ensure memory consistency. By iterative it means that pre-

copying occurs in several rounds and the data to be transmitted

during a round are the dirty pages generated in the previous

round. The pre-copying phase terminates (1) if the memory

dirtying rate exceeds the memory transmission rate; or (2) if

the remaining dirty memory becomes smaller

Than a pre-defined threshold value; or (3) if the number of

iterations exceeds a given value; or (4) the network traffic

exceeds a multiple of the VM memory size. After several

rounds of synchronization, a very short stop-and-copy phase is

performed to transmit the remaining dirty pages.

As the data transferred is relatively small, this mechanism

results in a nearly negligible best-case migration downtime.

We note that the performance of live VM migration is affected

by many factors. First of all, the size of VM memory has a




main effect on the total migration time and network traffic.

Secondly, the memory dirtying rate, which reflects the

memory access pattern of different applications, impacts the

number of iteration rounds and data transferred in each pre-

copying round, and hence indirectly affects the migration

latency and network traffic.

Thirdly, the network transmission rate together with the

configuration of migration algorithm is also crucial to

migration performance. However, migration performance

varies significantly depending on the workload

characterization even when other conditions remain the same.

For instance, migration of a VM running memory-intensive

applications would lead to more performance penalty in terms

of network traffic, migration downtime and latency [12].

Fig. 1. Live migration algorithm performs pre-copying in iterative rounds[12]

IV. WEIGHTED FAIR QUEUING METHOD[18]

WFQ1 is a data packet scheduling technique allowing different

scheduling priorities to statistically multiplexed data flows.

WFQ is a generalization of FQ2. Both in WFQ and FQ,

each data flow has a separate FIFO queue. In FQ, with a link

data rate of R, at any given time the N active data flows (the

ones with non-empty queues) are serviced simultaneously,

each at an average data rate of .

Since each data flow has its own queue, an ill-behaved flow

(who has sent larger packets or more packets per second than

the others since it became active) will only punish itself and

not other sessions.

As opposed to FQ, WFQ allows different sessions to have

different service shares. If N data flows currently are active,

with weights 1+ 2+… N , data flow number i will achieve

an average data rate of

It can be proven [15] that when using a network with WFQ

switches and a data flow that is a leaky bucket constrained, an

end-to-end delay bound can be guaranteed. By regulating the

WFQ weights dynamically, WFQ can be utilized for

controlling the quality of service, for example to achieve a

guaranteed data rate .

1 Weighted Fair Queuing 2 Fair Queuing

Fig. 2. Example of weighted fair queuing[15]

V. PROPOSED METHOD FOR VIRTUAL CLUSTERS

MIGRATION

In a distributed Cloud environment of large-scale, services can

to be divided into smaller parts, allowing services to provide

through a set of virtual machines (clusters).

In this dynamic environment, services have moved from a

place to another. So if several services may need to migrated

simultaneously, all the services or a set of virtual machines

can't be migrated simultaneously, because of bandwidth

constraints.

In the proposed Method, We focus on the migration of

multiple virtual clusters.

In this algorithm, a weight is allocated to each virtual cluster,

which is obtained based on memory size of the virtual

machine, if is larger virtual machine's memory size, the more

weight will be allocated to the cluster elements.

This weight assigned to each cluster causes to obtain a

separate portion of the bandwidth. Because have different

sizes of virtual machines within each cluster, first select a

virtual machine that has the most weight in a cluster, is put

into a queue and will be migrated in its bandwidth.

Algorithm 1- Virtual Clusters Migration input:

MigrationList for each Cluster

LinkSpeed

Weight: for each Cluster

Curent_Time

VmMigration_Time

ClusterMigration_Time

FinishTime=0

TotalMigrationTime

While (MigrationList != Null)

for i:0 to MigrationList_Size

if VmMigration_Time is finished

List.add( ith VM in MigrationList)

List sort by Asc

for j:0 to list_size

List_VM[j] is migrated

VmMigration_Time:

Current_Time +( List_VM[j] / (Link_Speed* Weight j/ )

FinishTime+= VmMigration_Time

for k:0 to MigrationList_Size

ClusterMigration_Time: Sum(VmMigration_Time for each Cluster)

TotalMigrationTime: max(sum)

Fig. 3. Pseudo-code of the WFQ algorithm for virtual Clusters migration


http://en.wikipedia.org/wiki/Data_packet

http://en.wikipedia.org/wiki/Scheduling_(computing)

http://en.wikipedia.org/wiki/Statistical_multiplexing

http://en.wikipedia.org/wiki/Flow_(computer_networking)

http://en.wikipedia.org/wiki/FIFO

http://en.wikipedia.org/wiki/Weighted_fair_queuing#cite_note-1

http://en.wikipedia.org/wiki/Leaky_bucket

http://en.wikipedia.org/wiki/Quality_of_service

http://en.wikipedia.org/wiki/Fair_queuing



Then, are selected next virtual machine in the cluster and

placed in the queue and then to be migrated. Also obtained a

finish time for migration of each virtual machine, Due to the

completion of the migration time of virtual machines, is

selected from each cluster a virtual machine with round robin

and to be migrated in its bandwidth.

Figure 3 is an example of a weighted fair queuing has been

shown to migrated virtual machines.

Fig. 4. Example of weighted fair queuing for virtual machine migration

A. Simulation of the proposed method

To simulate the proposed algorithm, we have defined a

scenario for migrating virtual clusters, so that, is designed two

data center which is considered five physical machines and a

Broker who is responsible for resources allocating.

At the first data center is located 20 virtual machines and in

the second data center 10 Virtual Machine.

This scenario has been implemented in the Java programming

language and CloudSim simulation. The results of these

simulations will be explained. We have applied in our

proposed method of weighted fair queuing algorithms.

For example, in our sample are ready to migrate three clusters

of virtual machines there are four virtual machines per cluster

and is the weight of each cluster are included W_VC1 = 1,

W_VC2 = 1 and W_VC3 = 3 and share the bandwidth of each

cluster is included bw_VC1 = 20%, bw_VC2 = 20% and

bw_VC3 = 60% and figure 4 shows the changes the migration

time for these three clusters.

The memory size of the virtual machines in the clusters varies

from 128 MB to 1024 MB. "Xen" is hypervisor each

Machines, "Linux" is OS 3 each Physical Machine. In this

example is memory size of the virtual machines VMSize_VC1

= 128, VMSize_VC2 = 128 and VMSize_VC3 = 256.

3 Operating System

Fig. 5. Total migration time in each cluster in the proposed method

RELATED WORK

Cloud computing provides a way to maximize the capacity

and capabilities without investing in infrastructure. The main

purpose of applying the technique migration, load balancing,

fault tolerance, energy management and maintenance of

servers and its main function is to improve the service. To

implement this technique has been proposed so many

methods.

Pre-copy technique [16, 17] is the classic mechanism to

implement the live migration in different hypervisors.

Ye et al [19] proposed a framework for migrating Virtual

clusters. They evaluated the performance and overhead of

virtual clusters live migration.

In [20] is used the post-copy technique to virtual machine

migration. In the basic approach, post-copy first suspends the

migrating VM at the source node, copies minimal processor

state to the target node, resumes the virtual machine, and

begins fetching memory pages over the network from the

source. The manner in which pages are fetched gives rise to

different variants of post-copy, each of which provides

incremental improvements. The result of this research is to

reduce the number of pages transmitted and total migration

time compared to the Pre-Copy technique.

CONCLUSION

The benefits of virtual machines clustering consist of powerful

processing, increase efficiency, reduce response time, simplify

the migration process and performance improvement and etc.

In this paper, we have proposed a weighted fair queuing

algorithm for migration of virtual clusters. In this algorithm,

for each cluster is considered a weight. The selection of the

virtual machine from each cluster is performed according to its

weight. Finally, these machines are placed in a queue and they

are migrated according to bandwidth, that this method makes

the increase efficiency and reduce the total migration time.




REFERENCES

[1] P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho, R.

N I. P A. W “X of ” P ACM Symposium on Operating Systems Principles, p. 177, 2003.

[2] R. Buyya, C. Yeo, S.Venugopal, J.Broberg., I.Brandic, “Cloud computing and emerging IT platforms” Future Generation Computer Systems 25 pp.599–616, 2003.

[3] M. Armbrust, A. Fox, R. Griffith, A. Joseph, R. Katz, A. Konwinski, G. Lee, D. Patterson, A. Rabkin, I. Stoica, . “A ” C ACM, vol. 53, no. 4, pp. 50–58, 2010.

[4] C. W “M VM ESX ” ACM SIGOPS Operating Systems Review, vol. 36, no. SI, p. 194, 2002.

[5] E.P.Zaw, N.L.Thein, “I L VM M U LRU S T A ”, International journal of Compute Science and Telecommunications, Volume 3, Issue 3, 2012.

[6] M. N B. L G. H “F migration for ” Proceedings of the annual conference on USENIX Annual Technical Conference, p. 25, 2005.

[7] K. Y D. H X. J H. C S. W “V machine based energy-efficient data center architecture for cloud computing: a ” Proceedings of the 2010 IEEE/ACM International Conference on Green Computing and Communications, pp. 171–178, 2010.

[8] K. Y X. J S. C D. H B. W “A and modeling the performance in Xen-based virtual cluster ” 2010 12th IEEE International Conference on High Performance Computing and Communications (HPCC), pp. 273–280, 2010.

[9] M. F. M V. U O. K J. X “V for high- ” SIGOPS O . Syst. Rev., vol. 40, no. 2, pp. 8–11, 2006.

[10] Y.W M.Z ”P M V M L M ” International Conference on Cloud Computing (CLOUD), IEEE , PP. 492-499, 2011.

[11] S.Ak R.S A.R A. .M A.H “Predicting The Performance Of Vir M M ” in Modeling, Analysis & Simulation of Computer and Telecommunication Systems (MASCOTS), IEEE International Symposium on , pp. 37 – 46, , 2010.

[12] H.Liu, H.Jin, C-Z.Xu, X.Liao, “Performance and energy modeling for live migration of virtual ”, Springer Science+Business Media, LLC 2011.

[13] A.Agarwal, S.Raina, Live Migration of Virtual Machines in Cloud, International Journal of Scientific and Research Publications, Vol.2, Issue 6, June 2012.

[14] D. S A. V “L -rate servers: a general model for ” IEEE/ACM T Networking, 1998.

[15] P. Bertsekas ” T B r and Queuing in a QOS Environment” Department of Electrical Engineering M.I.T, OPNET Technologises, 2002.

[16] C. Clark, K. Fraser, S. Hand, J.G Hansen, E. Jul, C. Limpach, I. Pratt, A. Warfield, “Live migration of virtual machines. In:Proceedings of Second S N k S D I (NSDI’05) ”, pp. 273–286, 2005.

[17] M. N B.H. L G. H “Fast transparent migration for virtual machines. In: Proceedings of USENIX Annual Technical Conference (USENIX’05)” Anaheim, California, USA, pp. 391–394, 2005.

[18] http://en.wikipedia.org/wiki/Weighted_fair_queuing

[19] K. Ye, x. Jiang, R.Ma, F.Yan, "VC-Migration: Live Migration of Virtual Clusters in the Cloud", ACM/IEEE 13th International Conference on Grid Computing, PP. 209-218, 2012.

[20] M.R.Hines, U.Eshpande, K.Gopalan, "Post-Copy Live Migration of Virtualization of Virtual Machine" International Conference on Virtual Execution Environment (VEE), pp. 14-26, March 2009.


http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=5581171



http://en.wikipedia.org/wiki/Weighted_fair_queuing

Fisher’s Linear Discriminant and Echo State Neural Networks for Identification of Emotions

Devi Arumugam and Purushothaman S.,

1Devi Arumugam Research Scholar, Department of Computer Science,

Mother Teresa Women’s University, Kodaikanal, India-624102.

2Dr. S. Purushothaman Professor, PET Engineering College,

Vallioor, India-627117

Abstract— Identifying the emotions from facial expression is a fundamental and critical task in human-computer vision. Here expressions like anger, happy, fear, sad, surprise and disgust are identified by Echo State Neural Network. Based on a threshold, the presence of an expression is concluded followed by separation of expression. In each frame, complete face is extracted. The complete face is from top of head to bottom of chin and left ear to right ear. Features are extracted from a face using Fisher’s Linear Discriminant function. The features are extracted from a face is considered as a pattern. If 20 frames belonging to a video are considered, then 20 patterns are created. All 20 patterns are labeled as (1/2/3/4/5/6) according to the labelling decided. The labelling is done as anger=1, fear=2, happy=3, sad=4, surprise=5 and disgust=6. If 20 frames from each video is obtained then number of patterns available for training the proposed Echo State neural Networks are 6 videos x 20 frames= 120 frames. Hence, 120 patterns are formed which are used for training ESNN to obtain final weights. This process is called during the testing of ESNN. In testing of ESNN, FLD features are presented to the input layer of ESNN. The output obtained in the output layer of ANN is compared with threshold to decide the type of expression. For ESNN, the expression identification is highest.

Keywords- Video frames; Facial tracking; Eigen Value and eigen vector; Fisher’s Linear Discriminant (FLD); Echo State Neural Network (ESNN);

I. INTRODUCTION The development of modern human computer interface

system requires computers to closely understand the biometrics of human system. Some of the biometrics is face, iris, fingerprint, GAIT. Biometrics through face, eyes place an important role in activating icons on the TV screen, computer screen. Hence, a technique of detection of expressions from face helps in various applications.

From the video images acquired from built-in cameras, and from speech waveforms collected from on-board microphones, this information can be used to teach computers to recognize human emotions. Computers may learn to recognize gestures, postures, facial expressions, eye contact.

The term “Emotional expression” means any outward expression that arises as a response to some stimulus event.

II. RESEARCH WORK Emotion recognition through the computer-based of facial

expression has been an active and interesting area of research in the literature survey for a long time.

Paul Ekman 1978, as reported by Ekman, anger and disgust are commonly confused in judgment studies. Also, fear and surprise are commonly confused. Because of sharing the similar facial actions, these confusions are occurred.

Bassili, 1979, compare several facial expression recognition algorithms. The author states that these algorithms perform well compared to trained human recognition of about 87%.

Paul Ekman 1994, and his colleagues have performed wide studies of human facial expressions. They found proof to support universality in facial expressions. These “universal facial expressions” are happiness, sadness, anger, fear, surprise, and disgust. They studied facial expressions in various cultures, with preliterate cultures, and found much unity in the expression and recognition of emotions on the face.

Essa, 1997, to extract motion blobs from image sequences, the spatial and temporal filtering together with thresholding is used. Turk 1991, to detect the presence of a face, the blobs are then evaluated using the eigen faces method Essa, 1997, extended their face detection approach to extract the positions of prominent facial features using Eigen features.

Chaudhury et al., 2003, described that instead of a fixed threshold value to initialize the face tracker, used two face probability maps. One used for frontal views and another one for profiles.

Boccignone et al., 2005, Li et al., 2006, proposed that before tracking a face, one should choose the features to track. The exploitation of color is one of the common choices because it is invariant to facial expressions, scale, and poses changes.

Arnaud et al., 2005; Zhiwei Zhu et al., 2005; Yan tong et al., 2007, described facial features extraction from eyes, nose and mouth.



Fig. 1 Sample frames for facial expressions extracted from video

III. WORKING DETAILS OF PROPOSED ARCHITECTURE

A. Emotion Data Simulation Fig 1 shows some of the sample frames for facial

expressions which were extracted from the video. The person is asked to express his/her feelings to different situations. The possible expressions are considered for this research work. The emotion expression for anger, disgust, fear, happy, sad and surprise obtained from videos. The videos have been shot on a female who had expressed reactions through her face when particular statements are read aloud.

For simulation of facial expression, Video has been shot for a female who was asked to react to various statements. Sample statements are “I will give you a new two wheeler”, “hey, you got selected”. The happiness expressed by the female for the two different statements will not be the same even through the female reacted with happiness. Hence, the expression happiness is expressed through face, eyebrows, rising of forehead, opening of mouth, rising of shoulders and many more. It would be a complicated process to combine all the movements of a body just for happiness. Hence, we have limited our attention only to facial movements in identifying expressions like Anger, happy, Fear, Sad, Surprise and Disgust

as these 7 categories are mostly used by the earlier researchers in their contributions.

B. Facial Tracking Tracking the presence of face in each frame is done for

subsequent processing. For each tracked face, three steps are involved: Initialization, tracking, stopping.

C. Facial Feature Representation using Fisher’s Linear Discriminant Function Emotional facial feature extraction is the process of

converting a face image into a feature vectors. The feature vector should represent a face. This vector is used as the basis for emotional expression classification. This vector for the emotional expression recognition must have all the essential features for the classification.

Foley, 1972, had discussed the method of considering the number of patterns and feature size. Siedlecki and Skalansky, 1988, have given an overview of mapping techniques. Fisher, 1936, has developed a linear classification algorithm. The fisher’s criterion is given by

(1) J() =TSb/TSW



(2)

(3)

Where

Sb is the between class matrix, and Sw is the within class matrix which is non-singular.

Equation (1) is the discriminant vector that maximizes J is denoted by ϕ1. Another discriminant vector ϕ2 is obtained by using the same criterion. The vector ϕ1 is found as a solution of the Eigen value problem. The vector ϕ2 should also satisfy the equation given by

(4)

Equation (4) indicates that the solution obtained is geometrically independent. The Discriminant vector ϕ2 is found as a solution of the Eigen value problem, which is given by below equation.

(5)

Where

λm2 is the greatest non-zero eigen value of QpSbSw-1 and Qp is

the projection matrix given by below equation.

(6)

Where, I is an identity matrix. Sw should be non-singular. Then the values of ϕ1 and ϕ2 are obtained.

Fig. 2 Plot of discriminant vectors for all 6 expressions

Fig 2 shows plots of 6 different expressions based on the FLD output. In this figure, the expressions happy and sad are

scattered and not cluttered. The other expressions are cluttered.

D. Training Artificial Neural Networks for identifying emotions

Step 1: Input a video. Step 2: Frames are extracted from a video. Important features are extracted from successive frames that belong to one second. Step 3: Input each frame to fisher’s linear discriminant function and obtain features. Step 4: The features are trained using the proposed ESNN and final weights are stored in a database.

E. Testing Artificial Neural Networks for identifying emotions

Step 5: In the testing process, Step 2 to step 4 are adopted. The extracted features are processed with final weights of the ESNN, to get an output in the output layer of the ESNN. Step 6: The output is compared with a threshold value, to decide the category to which the particular emotion the facial expression belongs.

IV. ECHO STATE NEURAL NETWORK FOR EMOTION FACIAL EXPRESSION IDENTIFICATION

A recurrent neural network has been proposed for emotion facial expression identification. The echo state condition is defined in terms of the spectral radius (the largest among the absolute values of the eigen values of a matrix, denoted by (|| ||) of the reservoir’s weight matrix (|| W ||<1). The recurrent network is a reservoir of highly interconnected dynamical components, states of which are called Echo states. The memory less linear readout is trained to produce the output. The topology of ESNN consists of M input units, N internal PEs, and L output units.

The value of the input unit at time n is u(n) = [u1(n), u2(n), . , , uM(n)]T, the internal units are x(n) = [x1(n), x2(n), . . . , xN(n)]T, and output units are y(n) = [y1(n), y2(n), . . . , yL(n)]T. The connection weights are given as follows: An N x M weight matrix Wback=Wij

back for connections between the input and the internal PEs, An N × N matrix Win = Wij

in for connections between the internal PEs, An L × N matrix Wout=Wij out for connections from PEs to the output units and An N × L matrix Wback=Wij back for the connections that project back from the output to the internal PEs. The activation of the internal PEs (echo state) is updated according to

(7)

In Equation (7), f = (f1, f2. . . fN) are the internal PEs’ activation functions. Here, all fi’s are hyperbolic tangent functions ex-e-x/ex+e-x. The output from the network is computed according to the below equation. (8)

Sb= ∑ P(ωi)(m1-m0)(m1-m0)T

Sw= ∑ P(ωi) E[(xi-mi)(xi-mi)T/ωi]

2T1= 0.0

QpSb2= λm2Sw2

1 1T Sw-1

Q = I 1T Sw-1 1

X(n+1) = f(Win u(n+1)+Wx(n)+Wbacky(n))

Y (n+1) = fout (Woutx(n+1))



In Equation (8), fout = (f1out, f2

out,…, fLout) are the output unit’s

nonlinear functions (Purushothaman et al, ).

A. Training the ESNN A state vector is initialized with zero. The length of

the state vector is equivalent to number of nodes or reservoirs in the hidden layer of the ESNN. The number of reservoirs is decided based on the minimum error obtained in the ESNN in estimation of the expression. The summation of (input pattern multiplied with initial weights between input and hidden layers, multiplication of initial state vector with the initial weights of the reservoir and multiplication of target value with the initial weights between hidden layer and output layer) is obtained. A new state vector is obtained by passing the summed value over an activation function which is the tanh function. Hence, 100 state vectors are obtained if there are 100 training patterns. An ESNN matrix is obtained whose size is number of training patterns (100) X number of reservoirs (21). This matrix is a rectangular matrix and hence a pseudo inverse of the ESNN matrix is found and multiplied with the target values to obtain final weights. The training phase of ESNN is described below: Step 1: Read emotion image Step 2: Decide the number of reservoirs = 21 or 22. Step 3: Decide the number of nodes in the input layer=2. Step 4: Decide the number of nodes in the output layer = number of target values =1. Step 5: Initialize state vector (number of reservoirs) = 0. Step 6: Initialize random weights between input layer (IL) and hidden layer (hL). Initialize weights between output layer (oL) and hidden layer (hL). Initialize weights in the reservoirs Step 7: Calculate state_vectornext = tanh ((ILhL)

weights*Inputpattern + (hL) weights* state vectorpresent+ (hLoL)weights * Targetpattern). Step 8: Calculate, a = Pseudo inverse (State vectors all patterns). Step 9: Calculate, Wout = a * T and store Wout for emotion facial expression classification.

B. Testing the ESNN A pattern with two FLD features obtained is

presented to the input layer of the ESNN. The summation of input pattern multiplied with final weights between input and hidden layers + multiplication of final state vector with the final weights of the reservoir + the final weights between hidden layer and output layer is obtained. The tanh (summation) is obtained and added with the already obtained value during training (pseudo inverse (state matrix) X target of all the patterns). The final value of the output of ESNN is compared with a threshold of (1/2/3/4/5/6/) to decide the type of emotional facial expression classification.

A state_vector is obtained by multiplying 2 FLD features with final weight matrices obtained during training. The obtained value is passed over tanh function. The resultant value is the output in the output layer. The testing phase of ESNN is described below:

Step 1: Adopt step 1 and step 2 mentioned in Training. Step 2: Calculate state vector = tanh ((ILhL)weights*Inputpattern + (hL)weights* state vectorpresent+ (hLoL)weights * Targetpattern). Step 3: Estimated output = state vector * Wout. Step 4: Based on output in step 4, decide the type of expression.

In order to obtain best estimation from ESNN, optimum values for different parameters of ESNN obtained. Deciding the number of reservoirs, range of initial weights in reservoir matrix and range of initial weights between reservoir and output layer gives a good emotional facial expression classification. The change of weight values and their impact in estimation of ESNN is presented when the weight normalization is done only between output layer and hidden layer (reservoirs). The error increases and decreases. Hence lesser weight range has to be used to obtain good estimation of type of emotional facial expression classification. The change of weight values and their impact in estimation of ESNN is presented when the weight normalization is done only between input layer and hidden layer (reservoirs). The error increases and decreases continuously. The weight should be in the range of 0.5-0.6 for increased accuracy of estimation of type of emotional facial expression classification. The change of weight values and their impact in estimation of ESNN is presented when the weight normalization is done only in reservoirs. The error increases and decreases continuously.

V. RESULTS AND DISCUSSION The recognition performance of the proposed algorithms for

classifying the type of emotions is disussed. Videos were taken for 6 expressions: Anger, happy, fear, sad, surprise, disgust. The frames are extracted from the videos and the extracted frame with some expression is classified. These frames are considered for feature extraction and classification of the emotion expression. Volume of data considered: 480 frames are considered from each category of expression.

The accuracy refers to how correctly; the proposed algorithms classify the facial emotion in a video. Different measures like precision and recall can be used to evaluate classification of emotion expressions. However, in this work the facial emotion expression classification accuracy is expressed as follows:

(9)

(10)

(11)

Where FP-Frame does not contain any expression, but algorithm says that the expression is present. FN-An expression is actually present in a frame. But the algorithm says that there is no expression. TP-True Positive, frame contains Facial Emotion Expressions (FEE). It is correctly classified.

Sensitivity = TP TP + FN

Specificity = FP FP + TN

Accuracy = TP + TN TP+TN+FP



TN-True Negative, frame classified as not containing FEE. Hence, the expression is not detected and classified. Detection refers to presence of an FEE. Classification refers to labelling 6 emotions.

Fig 3 ROC for ESNN

The Fig.3 presents ROC for the performance of ESNN in estimating “angry expression”. The ROC plot shows the points above the diagonal. The TPR is more and the FPR is less.

Figure 4 Accuracy for ESNN in identifying “Angry’ expression

The Fig.4 presents accuracy for ESNN. The accuracy is best for all the 20 videos.

Fig 5 Specificity for ESNN in identifying ‘Angry” expression

The Fig.5, Fig.6 presents specificity and sensitivity obtained for the ESNN. The specificity, sensitivity of ESNN is best for all the 20 videos.

V.CONCLUSION Three features were extracted using fisher’s linear

discriminant. These three features are found to be optimal to train ESNN. The ESNN provides high accuracy in recognizing emotion.

Fig. 6 Sensitivity for ESNN in identifying “Angry” expression

REFERENCES [1] Arnaud E., Fauvet B., Mémin É., and Bouthemy P., 2005, A robust and

automatic face tracker dedicated to broadcast videos, In Proceeding Of International Conference on Image Processing, pp.429-432.

[2] Boccignone G., Caggiano V., Fiore G. D., and Marcelli A., 2005, Probabilistic detection and tracking of faces in video, In Proceedings of International Conference on Image Analysis and Processing, pp.687-694.

[3] Bassili J.N., 1979, Emotion recognition: The role of facial movement and the relative importance of upper and lower areas of the face, Journal of Personality and Social Psychology, Vol.37, No.11, pp.2049-2058.

[4] Choudhury R., Schmid C., and Mikolajczyk K., 2003, Face detection and tracking in a video by propagating detection probabilities, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.25, No.10, pp.1215-1228.

[5] Essa I.A., and. Pentland A.P., 1997, Coding, analysis, interpretation and recognition of facial expressions, IEEE Transaction on Pattern Analysis and Machine Intelligence, Vol.19, No.7, pp.757- 763.

[6] Fisher R.A., 1936, The use of multiple measurement in taxonomic problems, Annals of Eugenics, Vol.7, pp.178-188.

[7] Foley D.H., 1972, Consideration of sample and feature size, IEEE Transactions on Information Theory, Vol.IT-18, No.5, pp.681-626.

[8] Li Y., Ai H., Huang C., and Lao S., 2006, Robust head tracking with particles based on multiple cues fusion, ECCV Workshop on HCI, pp.29-39.

[9] Paul Ekman., and Friesen., 1978, W.V. Facial Action Coding System: Investigator’s Guide, Consulting Psychologists Press.

[10] Paul Ekman., 1994, Strong evidence for universals in facial expressions: A reply to Russell’s mistaken critique, Psychological Bulletin, Vol.115, No.2, pp.268-287.

[11] Siedlecki W., Siedlecka K., and Skalansky J., 1988, An overview of mapping techniques for exploratory data analysis, Pattern Recognition, Vol.21, No.5, pp.411-429.

[12] Siedlecki W., Siedlecka K., and Skalansky J., 1988, Experiments on mapping techniques for exploratory pattern analysis, Pattern Recognition, Vol.21, No.5, pp.431-438.

[13] Yan tong., Yang Wang, Zhiwei Zhu, Qiang Ji, 2007, Robust facial feature tracking under varying face pose and facial expression, ELSEVIER, Pattern Recognition, Vol.40, No.11, pp.3195-3208.

[14] Zhiwei zhu, Qiang ji, 2005, Robust real-time eye detection and tracking under variable lighting conditions and various face orientations, ELSEVIER, Computer Vision and Image Understandings, Vol.98, No.1, pp 124-154.



A New Current-Mode Multifunction Inverse Filter Using CDBAs

Anisur Rehman Nasir Syed Naseem Ahmad

Dept. of Electronics and Communication Engg. Jamia Millia Islamia, New Delhi-110025, India

Abstract: A novel current-mode multifunction inverse filter configuration using current differencing buffered amplifiers (CDBAs) is presented. The proposed filter employs two CDBAs and passive components. The proposed circuit realizes inverse lowpass, inverse bandpass and inverse highpass filter functions with proper selection of admittances. The feasibility of the proposed multifunction inverse filter has been tested by simulation program. Simulation results agree well with the theoretical results. Keywords : CDBA, multifunction, inverse filter

1. INTRODUCTION

The design of inverse filter is useful in Communication and instrumentation engineering. These filters are used to reverse the distortion of signal incurred due to signal processing and transmission.The distorted signal is to be converted to the input signal. The inverse filtering is used to inverse the transfer characteristic of the original signal [1, 2].

Several continuous time analog inverse filters are available in literature [2-11].The most of the inverse filter circuits available in literature are voltage–mode circuit. The voltage mode inverse filters are realized generally by CFOAs, CCIIs and CDBAs [6,7,9,11]. In current-mode the inverse filters and allpass filter have been realized using FTFNs [2-5] and CDTA [8] respectively.

Leuciuc [2] proposed a general method for realizing inverse filter using nullors. B. Chipipop et al [3] and H. Y. Wang et al [4] proposed current-mode universal filters using FTFNs. M.T. Abuelmatti proposed current-mode inverse filter using FTFN [5]. S.S. Gupta et al [6] and H.Y. Wang et al [7] proposed voltage mode inverse filter configuration using CFOAs that realize inverse lowpass, inverse highpass and inverse bandpass filter from suitable choice of admittances. N.A. Shah et al [8] proposed inverse allpass filters using CDTAs. R. Pandey et al [12] proposed voltage-mode universal inverse filter using CDBAs which realizes all basic inverse filter functions. However to our knowledge, there are no current-mode inverse filter using CDBA. Therefore, in his communication an effort is made to realize current-

Dept. of Electronics and Communication Engg. Jamia Millia Islamia, New Delhi-110025, India

Mode multifunction inverse filter using CDBAs. The proposed circuit realizes all the basic filter functions in inverse mode i.e.inverse lowpass (ILP), inverse highpass (IHP) and inverse bandpass (IBP) by proper selection of types of admittances.

2. CIRCUIT DESCRIPTION

The current differencing buffered amplifier (CDBA) is recently introduced as an active element [10]. The CDBA is suitable for realization of current-mode continuous time filterfunctions because of several advantages like free from parasitic capacitances, differential nature at its input port, high slew rate and wide bandwidth.

The circuit symbol of CDBA is shown in Fig.1 and its port relations are given in equation.

, , and (1)

Fig.1 Circuit Symbol of CDBA

The proposed current-mode multifunction inverse filter circuit is shown in Fig.2. The routine analysis of circuit yields the current transfer functionsas follows

IOI

Y YY Y Y Y Y Y 2

where

n

w

z

ip

in

Vp

Vn

Vw

Vz

p Vw

Vz

CDBA

iw

ip



If admittances chosen areY sC G , and Y sCG , then

N s sC G sC G

3

Fig.2 Proposed Multifunction Inverse Filter

By proper selection of admittances forY ,Y ,Y , and Y as shown in Table-I, differentinverse filter functions can be realized.

Table I

Response ILP 0 0 IHP 0 0 IBP 0 0

The transfer function of the ILP, IHP and IBP can be expressed as

IOI

1G G s C C s C G C G G G⁄ 4

IOI

1s C C s C C s C G C G G G⁄ 5

IOI

1s C G s C C s C G C G G G⁄ 6

The natural angular frequency and the pole Q-factor of the filter are

7

1 8

The gain constants of ILP, IHP and IBP responses are given by

, ,

3. SENSITIVITY ANALYSIS

The passive sensitivities of and Q for the proposed current-mode inverse filtercan beexpressed as

12

12

12

It is observed that the passive sensitivities are lesser than unity in magnitude. Hence the performance of proposed current-mode multifunction inverse filters are not affected.

4. SIMULATION RESULT

The proposed current-mode multifunction inverse filter has been simulated with simulation software. The multifunction filter has been designed for fO =796.18 KHz and Q=1. The CDBAs have been realized withcommercially available AD844s. The equal values of passive components are used. The supply voltages are ±12V. All the resistors are taken as 10KΩ and capacitors as 20pF. The simulated frequency characteristics for inverse lowpass, inverse bandpass and inverse highpass filter functions are shown in Fig.4. The simulation results agree well with theoretical analysis of the filter.

p

n

z

w

n

p w

z

Iin

Y1

Y4 Y2

Y5

Y6

Y3

IO

CDBA1

CDBA2



(a) Inverse Lowpass Filter

(b) Inverse Bandpass Filter

(c) Inverse Highpass Filter

Fig.4 Frequency Response of InverseFilters

5. CONCLUSION

A new current-mode multifunction inverse filter using CDBAs has been presented. The proposed circuit uses two CDBAs and passive elements. The inverse lowpass, inverse bandpass and inverse highpass filter functions are realized by proper selection of passive elements. The

simulation results are in agreement with theoretical analysis.

REFERENCES

[1]J. K. .Tugnait, “Identification and deconvolution of multichannel linear non-Gaussian processes using higher order statistics and inverse filter criteria,” IEEE Transactions on Signal Processing, vol.45,no.3,p.658–672,1997.

[2]A.Leuciuc, “Using nullors for realisation of inverse transfer functions and characteristics,” ElectronicsLetters,vol. 33, no. 11, p. 949-951,1997

[3]B. Chipipop and W. Surakampontorn, “Realisation of current-mode FTFN-based inverse filter,” Electronics Letters, vol. 35, no. 9, , p. 690-692. 1999

[4] H. Y. Wang and C. T. Lee, “Using nullors for realisation of current-mode FTFN-based inverse filters,” Electronics Letters, vol. 35, no. 22, , p. 1889-1890. 1999

[5] M. T. Abuelma’atti, “Identification of cascadable current-mode filters and inverse filters using single FTFN,” Frequenz, Vol. 54, No. 11, , p. 284-289. 2000

[6] S. S. Gupta, D. R. Bhaskar and R. Senani, “New analogue inverse filters realised with current-feedback op- amp ”, International Journal of Electronics, vol. 98, no. 8, p. 1103–1113, 2011

[7]H.Y. Wang, S.H.Chang, T. Y. Yang, and P.Y. Tsai, “A Novel multifunction CFOA-based inverse filter”, Circuits and Systems, 2, 14-17 , 2011,

[8]N. A. Shah, M. Quadri, and S. Z. Iqbal,“High output impedance current-mode allpass inverse filter using CDTA,” Indian Journal of Pure and Applied Physics,vol.46,no.12,p.893–896,2008.

[9]N. A. Shah and M. F. Rather, “Realization of voltage-mode CCII based allpass filter and its inverse inversion”, India J. of Pure and Applied Physics, 44, 3, 269-271, 2006

[10]C. Acar and S. Ozoguz, “A new versatile building block: current differencing buffered amplifier suitable for analog signal processing filters,” Microelectronics Journal,vol.30,no.2,p.157–160, 1999.

[11]R. Pandey, N. Pandey,T. Negi and V.Garg, “CDBA based universal inverse filter”, ISRN ElectronicsVolume2013, 2013

T

Frequency (Hz)10k 100k 1M 10M

Gai

n (d

B)

0.00

10.00

20.00

30.00

T


Gai

n (d

B)

0.00

10.00

20.00

30.00

40.00

T


Gai

n (d

B)

-20.00

0.00

20.00

40.00

60.00

80.00




Assessment of Customer Credit through Combined Clustering of Artificial Neural

Networks, Genetics Algorithm and Bayesian Probabilities

Reza Mortezapour Department of Electronic And Computer

Islamic Azad University Zanjan, Iran

Mehdi Afzali Department of Electronic And Computer

Islamic Azad University Zanjan, Iran

Abstract—Today, with respect to the increasing growth of demand to get credit from the customers of banks and finance and credit institutions, using an effective and efficient method to decrease the risk of non-repayment of credit given is very necessary. Assessment of customers' credit is one of the most important and the most essential duties of banks and institutions, and if an error occurs in this field, it would leads to the great losses for banks and institutions. Thus, using the predicting computer systems has been significantly progressed in recent decades. The data that are provided to the credit institutions' managers help them to make a straight decision for giving the credit or not-giving it. In this paper, we will assess the customer credit through a combined classification using artificial neural networks, genetics algorithm and Bayesian probabilities simultaneously, and the results obtained from three methods mentioned above would be used to achieve an appropriate and final result. We use the K_folds cross validation test in order to assess the method and finally, we compare the proposed method with the methods such as Clustering-Launched Classification (CLC), Support Vector Machine (SVM) as well as GA+SVM where the genetics algorithm has been used to improve them.

Keywords- Data classification; Combined Clustring; Artificial Neural Networks; Genetics Algorithm; Bayyesian Probabilities.

I. INTRODUCTION

Today, with respect to the development of database systems and large amount of data saved in these systems, we need an instrument to process the data saved and to provide the users with the information resulted from the process. Data analysis is one of the most important methods that it provides the users and the analysts with some useful models of data with at least intervention of known users in order to make critical and important decisions of organizations according to them. Classification is one of the most common duties of data analysis. In fact, classification has been defined as evaluation of the characteristics of data set and then to allocate them to a set of groups predefined. Data analysis can be used to create a model or a view of a group based on data characteristics by

using historical data. Then, we can use the predefined model in order to classify new data sets. Also, we can use it for the future predictions by determining a view that is correspondent with it. Commercial issues such as regression analysis, risk management and case targeting are involved in the classification. In order to overcome the financial problems of credit, organizations and institutions have considered several sections as the credits management. The purpose of the company credits management is to determine policies and to observe strategies that are correspondent with the company's functional aspect in terms of risk and efficiency. If the customers observe the previsions of credit contracts and pay the cash of goods purchased on credit, the company efficiency would be increased. Risk or hazard is a probability that the company credit not be receipted or in order to receipt previous credits, the company would be incurred additional costs.

II. PREVIOUS RESEARCHES

In the past, many researchers provided traditional statistical methods to credit accounts by using Linear Discriminant Analysis (IDA) and Logistic Regression and it has been used two common statistical methods in the structure of credit rating models. Nevertheless, Krles, Prakash, Reichert and Wagner cho suggested that usually because of considering the classification nature of credit data, IDA is be needed and this fact has been challenged that it seems unlikely to be the covariance matrix of bad and good credit groups. In addition to IDA method, logistic regression is usually another method for the credit rating. Logistic regression is a model that would be used to predict the probability of an event occurring. This method allows us to use different predictor variables which may be numerical or classified. Basically, the logistic regression model has initially been used as a method to predict binary outcomes. The logistic regression model doesn't need to the normal multi-variables hypothesis, but it depends on various access of perfect linear relationship between the independent variables for powering the logistic function.



Thomas and West showed that both logistic regression and IDA methods have this tendency to have a linear basic relationship between the variables and thus, it have been reported that it hasn't enough accuracy for rating credit. Recently, new database search methods can be used to make the credit rating models. Desai, Crook and Over Street used Neural Networks (NW), Logistic Regression and IDA to make the credit rating models. The results indicated that Neural Networks (NN) is showing the expectancy, whereas assessment of bad loans percent performance has been carefully classified. Nevertheless, IRA is as good as NN, whereas the criterion is percent performance of good and bad loans that it has been carefully classified. West compared the accuracy of credit points of five neural models and he reported that the hybrid structure of neural network models must be considered for the applicants of credit points. In addition, Hu, Kuo and Ho suggested the two-step search method that it uses the self-organizing plan to determine the number of clusters and then the algorithm of K methods would be used to classify the clusters samples. In this study, multiple using of clustering methods and neural networks would be affected by design of credit points' model. Malhotra and Malhotra compared the operation of Artificial Neural-Fuzzy Interface System (ANFIS) and different models of Discriminant Analysis to potential defaults screening of customer's loans. This result reported that in order to identify the bad credits demand, ANFIS is better than different methods of discriminant analysis. In recent years, the Support Vector Machine (SVM) was introduced to investigate the problems of classification (the demand for a new classification method). Many researchers used SVM method to rate the credit and to predict the financial risks, and the results obtained were promising. In addition, Hung, Chen and Wang chose three strategies to make the hybrid models of SVM-based credit points and to investigate the customer's credits points through the characteristics of customer input.

III. THE PROPOSED METHOD

In this study, a method has been proposed to assess the customer's credit that it uses three classifiers including Artificial Neural Networks, Genetics Algorithm and Bayesian Classifier, and then it extracts the final result obtained from above methods by a mechanism.

Fig. (1) shows the workflow of the proposed method. In the following sections, we will describe each section. Of course, due to the clearness of Bayesian classifier, we will not describe this issue and will express the experiments and the results at the end of this paper.

Figure.1 Workflow of the proposed method.

A. Artificial Neural Network After back-propagation training, multi-layer perceptron

networks are usually considered as a sample of standard networks for modeling the prediction and the classification: selecting an optimal MLP architecture is one of the areas that has been studied.

Method of function of multi-layer perceptron neural network with back-propagation training essentially consists of two main paths. The first path is called forward path where the input vector is applied to MLP network and its effects would be propagated from the middle layers to the output layers. The output vector formed in the output layer is true response of MLP network. In this path, the network parameters would be considered constant and invariable. The second path is called backward path. Unlike the forward path, the parameters of MLP network would be changed and adjusted in this path. This adjustment would be done according to the error correction code. The error signal is formed in the output layer of the network. The error vector is defined as the difference between the optimal and the true response of the network. After calculating in the backward path of output layer and through the networks layers, the amount of error would be distributed in the entire network. Since the recent distribution has been done in a path contrary to the weight communications of synapses, the term back error propagation has been selected in order to describe the behavioral modification of the network. The network parameters would be adjusted in such a way that the true response of the network is as more optimal as possible. After making a multi-layer perceptron neural network through the back-propagation training, some decisions must be made that they have been shown in the following.

1) Activation Function of Neurons In a typical application that several inputs have been coded as 0-1, the neuron outputs are 0-1 with the annular activation functions and they are approximately -1 and +1 with the hyperbolic tangent activation function. In this condition, the hyperbolic tangent is the best option. We used the annular and the hyperbolic functions in German and Australian datasets, respectively.



2) Learning Rules The main rule of learning that has been provided by Rommel Hart is called Delta Rule. This rule is a way that the artificial neural network learns from his mistakes, and common processing for this learning consists of three actions.

a) Outputs Calculations b) Comparing Outputs with Optimal Solutions c) Adjusting the Weights and Bias Values and

Repeating the Processing

Learning usually begins by randomizing the weights and the bias values. The difference between the real output and the optimal output is called delta. It is better to minimize the delta. Decreasing the delta would be done through the weights and bias values, which this kind of learning is so-called Supervised Learning too. In this kind of learning, the important and the effective issue in the final results is the number of repetitions. Table (1) shows different repetitions of the network and validities obtained to illustrate this issue. There is another kind of learning that is called Unsupervised Learning. In this kind of learning, just the network input would be motivated and the network is self-organizing. In this way, the network would be structuralized internally. Thus, each hidden processing element responses strategically to a different set of input stimuli and it is not be used any knowledge to classify the outputs. In this way, there is this probability that the network can or can't produce a meaningful issue to everyone who is learning. Some examples of this learning are Correspondence Theory and Kohonen's Self-organizing Maps.

Number of occurrences Australian German

300 75.07 73.4

500 67.54 73.4

1000 70.14 73.4

2000 74.93 78.6

2500 71.01 78.2

3000 75.65 81.4

4000 78.12 80.2

Table 1: Number of iterations in training and the accuracy obtained.

3) Learning Rate

Learning rate is the last key that must be determined for decision-making. Most of the people used the learning rate in a way that they choose it with a large number close to 1. Optimal learning rates resulted from the smooth level of RMS error. If the graph of RMS error has the high increasing and decreasing variations in the output layer, it is clear that the learning rate utilized is not optimal and it should be decreased equally to all of the layers. We have set the learning rate for both datasets equal to 0.7.

B. Genetics Algorithm 1) Determining The From Of Solution In The Genetics

Algorithm In the standard form of the genetics algorithm, the solutions

are as the binary strings, but using this form for many problems leads to complicate the solutions and in many cases, providing the solution in this form will be impossible. Therefore, in the genetics algorithm applications for the optimization problems, instead of using the complicated binary strings, we used a solution form corresponded with the proposed problem. In this problem, we also used the solution form of the problem.

2) Method Of Determining Initial Population Of The Genetics Algorithm

In the standard genetics algorithm, the initial population would be achieved randomly. This method may be appropriate for unlimited problems. But in some other problems, the initial solutions can't be determined randomly, because there is no guarantee to exist the solutions. Thus, we have to select the initial population in a way that all of the solutions are justified. We also used the other methods' training data for the initial population.

3) Genetic Operations

We usually try to choose the operations in a way that the proportion rate of new responses (Children) is better than the parents. In the genetics algorithm utilized, we used two-point mixture operator as well as the mutation operator in the fields that were possible.

4) Recognition and Selection

So far, we have achieved three groups of responses by adjusting the parameters required for three methods used in the proposed method. In each method, graphs 1 and 2 of the validity obtained have been represented for German and Australian datasets. Now we extract the response. We test three methods for the final result. The first method is to use the majority voting, the second is to use a neural network and the third method is to use weighting for each of the methods. In this method, if we consider each of the methods as "V", we can use the following formula to extract the final result:

∑ (1)

where R is the result, n is the number of methods and V is the result of method obtained. Table 2 shows the results obtained from three methods mentioned above for two datasets.

IV. EXPERIMENTS

All of the results provided are resulted from running the programs on a system having characteristics such as Memory 3GB, Intel Pentium 2.2 GHZ and XP operating system. We used MATLAB and VB.Net 2008 programming language to implement the program. In order to certify this methods, we used the k_fold cross validation in the results provided, where k is equal to 10.



For all the tests, we used two datasets that their characteristics have been represented in Table 2.

Dataset Name

Record number

number of

numeric field

number of non-

numeric field

Record number of class

1

Record number of class

2 Australian

Credit Approval

690 6 8 307 383

German credit

dataset 1000 7 13 700 300

German credit

dataset-numeric

1000 24 0 700 300

Table. 2 Characteristics of data sets used in the paper

Figure. 1 Results of different methods on the German dataset

Figure. 2 Results of different methods on the Australian dataset

Since each algorithm had a different result compared to other algorithms, thus we have linked each of the algorithms' output to a voting system that has been shown in Fig. (1). Here, we considered three strategies to the voting system that Table 2 shows the results obtained from each of the strategies and their comparison to the previous three methods. We worked in a

method that in the first method, the majority voting was done; that is each of the algorithms has the same impact on the output. In the second method, the algorithms' outputs would be entered in a multi-layer perceptron neural network and with respect to the learning rate of 0.7 that is considered to the network, it represents the results of the output mentioned above, which it has an appreciable improvement compared with the majority voting. The third strategy is based on an appropriate weighting to each algorithm; that is with respect to the results obtained from each algorithm and its impact on the final result, we choose a weight according to it. In this method, we have set the weight of 0.5, 0.29 and 0.21 to neural network, Bayesian and genetics algorithm method, respectively.

V. RESULTS

In this system, regarding the turnovers conducted and their impact on the refund of the credit allocated to previous customers or all of the people that their information is available, we have evaluated the importance of each item and thus, we have omitted incorrect relationships and characteristics. In this paper, we have provided a comprehensive system to assess the customer's credit that it can significantly solves the problems of existing systems. This system can assess the credit and it can appropriately distinguish the credit decision-makings with a high accuracy despite registering incorrect information in data entry due to using different techniques and methods of data analysis. The proposed system has not environmental dependency; that is we can use this system in any environment due to the need for primary data. This system can provide different assessments for political and military applications in order to find the credit of proposed sections according to the activity has been asked.

German Australian The method used 83.3 78.2 Majority voting method

87 88.95 Artificial neural network method

90 84.7 based on the weighted voting method

84.80 86.52 Clc

73.70 80.43 Mysvm

77.92 86.90 GA+svm

Table. 2 Comparison of Methods

REFERENCES

[1] D Caudill,M., and C.Butler , Understanding Neural Networks: Computer

Explorations, vol.1 and 2 , Cambridge,MA:the MIT Press,1992.

[2] David Hand, Heikki Mannila , Padhraic Smyth. Principles of Data Mining. The MIT Press . 2001.

[3] J.Han, and M.Kamber, "Data Mining: Concepts and Techniques", San Diego Academic Press, 2001.

60

65

70

75

80

85

90

1 2 3 4 5 6 7 8 9 10

Accuracy Pe

rcen

t

Fold

Bayesian:

Neural Network:Genetic:

60

65

70

75

80

85

90

1 2 3 4 5 6 7 8 9 10

Accuracy Pe

rcen

t

Fold

Bayesian:

Neural Network:Genetic:



[4] Sameh M. Yamany, Kamal J. Khiani, Aly A. Farag,“Application of neural networks and genetic algorithms in the classification of endothelial cells”,Pattern Recognition Letters 181997.1205–1210.

[5] Gloria Curilem, Jorge Vergara, Gustavo Fuentealba, Gonzalo Acuña, Max Chacón, “Classification of seismic signals at Villarrica volcano (Chile) using neural networks and genetic algorithms”,Journal of Volcanology and Geothermal Research 180 (2009) 1–8.

[6] Chipman, H.A., George, E.I., McCulloch, R.E., 2010. “BART: Bayesian Additive Regression Trees”. Annals of Applied Statistics .

[7] Denison, D., Mallick, B., Smith, A., 1998. “ A bayesian CART algorithm”. Biometrika 363377.

[8] Robert a.marose,”a financial neural network application.” Ai expert, may 1990

[9] Antonio J. Tallón-Ballesteros, César Hervás-Martínez,“A two-stage algorithm in evolutionary product unit neural networks for classification”,Expert Systems with Applications 38 (2011) 743–754.

[10] Yi-Chung Hu, Jung-Fa Tsai,“Evaluating classification performances of single-layer perceptron with a Choquet fuzzy integral-based neuron”,Expert Systems with Applications 36 (2009) 1793–1800.

[11] Shu-Ting Luo, Bor-Wen Cheng, Chun-Hung Hsieh,“Prediction model building with clustering-launched classification and support vector machines in credit scoring”,Expert Systems with Applications 36 (2009) 7562–7566.

[12] Shu-Ting Luo, Bor-Wen Cheng, Chun-Hung Hsieh,“Prediction model building with clustering-launched classification and support vector machines in credit scoring”,Expert Systems with Applications 36 (2009) 7562–7566.

[13] M. H. Wang and C. P. Hung. “Extension Neural Network and Its applications.” Neural Networks, vol. 16, no. 5-6, pp. 779–784, 2003.

[14] Branke, J., 1995. “Evolutionary algorithms for neural network design and training”.Finland, pp.1–21.

[15] Xing Zhong, Gang Kou ; Yi Peng " A dynamic self-adoptive genetic algorithm for personal credit risk assessment " Information Sciences and Interaction Sciences (ICIS), 2010.

[16] Taremian, H.R. Naeini, M.P. " Hybrid Intelligent Decision Support System for credit risk assessment" Information Assurance and Security (IAS), 2011.

[17] Marikkannu, P. Shanmugapriya, K. " Classification of customer credit data for intelligent credit scoring system using fuzzy set and MC2 Domain driven approach" , Electronics Computer Technology (ICECT),2011.




A Cross Layer UDP-IP protocol for Efficient

Congestion Control in Wireless Networks

Uma S V, K S Gurumurthy

Department of ECE,

University Visvesvaraya College of Engineering,

Bangalore University

Bangalore, India

Uma S V1, K S Gurumurthy

2

Department of ECE, 1R N Shetty Institute of Technology,

2Reva Institute of

Technology and Management, Visvesvaraya Technological

University, Bangalore, India

Abstract—Unlike static wired networks, mobile wireless networks

present a big challenge to congestion and flow control algorithms

as wireless links are in a constant competition to access the

shared radio medium. The transport layer along with IP layer

plays a major role in Congestion control applications in all such

networks. In this research, a twofold approach is used for more

efficient Congestion Control. First, a Dual bit Congestion Control

Protocol (DBCC) that uses two ECN bits in the IP header of a

pair of packets as feedback is used. This approach differentiates

between the error and congestion-caused losses, and is therefore

capable of operating in all wireless environments including

encrypted wireless networks. Secondly, for better QoS and

fairshare of bandwidth in mobile multimedia wireless networks,

a combined mechanism, called the Proportional and Derivative

algorithm [PDA] is proposed at the transport layer for UDP

traffic congestion control. This approach relies on the buffer

occupancy to compute the supported rate by a router on the

connection path, carries back this information to the traffic

source to adapt its actual transmission rate to the network

conditions. The PDA algorithm can be implemented at the

transport layer of the base station in order to ensure a fair share

of the 802.11 bandwidth between the different UDP-based flows.

We demonstrate the performance improvements of the cross

layer approach as compared to DPCP and VCP through

simulation and also the effectiveness of the combined strategy in

reducing Network Congestion.

Keywords—congestion; explicit congestion bits [ECN];

transport layer; Internet Protocol [IP]; transmission rate;

I. INTRODUCTION

Mobile wireless networks present a big challenge to

congestion and flow control algorithms as wireless links are in

a constant competition to access the shared radio medium and

are also affected severely by random losses. Furthermore,

CSMA/CA-based wireless links suffer dramatically from

neighborhood interferences, where packet transmission

decisions are sensibly affected by carrier sensing within the

interference range as well as the use of the RTS/CTS

mechanism. Besides, the presence of random losses due to the

wireless transmission properties is a non-negligible

phenomenon that worsens the performances of such networks.

All these factors contribute in the well-known performance

degradation of wireless wide-spreading networks. Therefore,

congestion control has to be considered in a different manner

compared to wired networks, and should be intensively

investigated.

The issue of Congestion control in wireless networks is

often dealt with two prominent techniques. First are the

Explicit congestion control schemes, where routers play an

important role, since they are well located to react to a

congestion state. When congestion occurs, they explicitly

inform the end hosts of this state by explicit control messages.

Feedback control information can be binary or explicit. One

such scheme is the Explicit Congestion Notification (ECN),

where each router marks a passing IP packet's header when an

incipient congestion is detected. The end hosts react to an

ECN-marked packet by reducing their transmission rates. A

second approach is derived from ATM forum’s rate-based

congestion control algorithms. In these schemes, the routers

explicitly determine the permissible throughput of the

bottlenecks and assign to each flow its fairshare according to

the available bandwidth.

In this work, a cross layer approach involving marking the

IP packet headers efficiently for congestion notification with

differentiation of the type of losses and then using a new

algorithm for allotting of fairshare bandwidth among the

competing UDP flows is proposed. In the first part of this

paper, we propose a new congestion control protocol, Dual bit

Congestion Control Protocol (DBCC) with two new schemes:

i) A novel distributed scheme that allows for operation within

wireless encrypted networks, and ii) A new heuristic loss

differentiating scheme that can distinguish between error

caused loss and congestion-caused loss. In DBCC, a

congestion level is carried by a chain of two packets and each

packet provides two bits out of four bits of information

associated with a congestion level. The routers compute and

distribute a congestion signal into two packets. The congestion

level can be specified by concatenating a group of two ECN

bits together from a pair of packets at an end node.

Incorporated with a novel heuristic algorithm, DBCC can

appropriately react to congestion caused loss while avoiding

unnecessary reductions of the sending window sizes in

response to error-caused loss.




In the second part of this paper a new router based

congestion control algorithm, the Proportional and Derivative

algorithm (PDA) is proposed. This PDA monitors the level of

network congestion through the occupancy of the buffer which

is maintained with the control target. Based on the difference

between the present and target occupancy, the PDA controller

associated with each link computes periodically a fair rate)

and forwards the result to the next gateways till the

destination. Using a feedback scheme, the destination supplies

the source with the minimal received fairshare. In their turn,

sources have to adapt their transmission rate according to

received fair rate.

We show in this paper through simulation results that the

Combined DBCC and PDA algorithms guarantee efficiency

and fairness simultaneously reducing the Network congestion

effectively. Besides, through analysis, choice of control gain

leading to system stability in term of traffic transmission rate

has been made. The paper is structured as follows: Section II

presents the related work, Section III discusses the proposed

cross layer approach for congestion control. In Section IV, we

present the implementation and simulation results. Section V

concludes the paper.

II. RELATED WORK

The two issues addressed in this work have been addressed

individually in the past. Over the last few years, abundant

techniques have been developed to improve the efficiency and

fairness of TCP. A network explicit feedback mechanism

based on link throughput measurement was developed in [1].

Another explicit call admission control scheme, called

EXACT is proposed in [2]. If the proposed explicit congestion

control schemes succeed to reach efficiency and fairness, none

of them is dealing with the stability criterion, which is one of

the most important issues in highly dynamic wireless

networks.

Examples include the works of [3], [4], [5] using

algorithms to adaptively adjust the sending window size, and

[6], [7], [8], [9] employing alternative congestion signals.

However, due to their integrated controller design, these

techniques often fail to achieve both efficiency and fairness

[10]. By decoupling efficiency control from fairness, eXplicit

Congestion-control Protocol (XCP) [11] and Variable-

structure Congestion-control Protocol (VCP) [12] can achieve

high utilization, low persistent queue length, insignificant

packet loss rate, and sound fairness depending on the

heterogeneity characteristics of a network. While XCP

requires the use of a large number of IP packet header bits to

relay congestion information thereby introducing significant

deployment obstacles, VCP only uses the two existing ECN

bits in the IP header to encapsulate three congestion levels

hence presenting a more practical alternative of deployment

than XCP. However, VCP can only deliver limited feedback to

end hosts since two bits can at most represent four levels of

congestion. In order to avoid sudden bursts, VCP has to

control the growth of transmission rates by setting artificial

bounds. This yields slow convergence speeds and high

transition times. Moreover, due to the use of fixed parameters

for fairness control, VCP exhibits poor fairness characteristics

in high delay networks. Very recently, several works have

attempted addressing the problem associated with VCP

limitations by increasing the amount of feedback. While the

work in MLCP [13] proposes using 3 bits to represent the

Load Factor (LF), the UNO framework [14] proposes another

alternative to increase the amount of feedback by passively

utilizing information in IP Identification (IPID) field. In

contrast, DPCP [15] proposes a distributed framework that

allows for using no more than 2 ECN-bits to deliver a 4-bit

representation of the LF. That said, DPCP needs to access

partial information in the TCP header in order to be able to

efficiently distribute and reassemble the LF. However, in

encrypted networks protected by IPSec, TCP header

information is lost when crossing encryption boundaries.

Thus, DPCP cannot operate in such encrypted networks.

Furthermore, wireless networks are characterized by fading

related error-caused loss in addition to queuing related

congestion-caused loss. Experiments have shown that the

performance of any congestion control protocols relies on

appropriate reaction to loss according to its source. Like VCP,

DPCP reacts to loss without differentiating between the

sources of loss and thus performs inefficiently over wireless

networks.

Considering the issue of fair share, it has been proven in

[16] and [17] that TFRC does not always meet its fair share

when the network conditions are dynamic and may present

TCP-unfriendliness behavior. In [18], the TFRC performance

degradation in wireless environment is highlighted and found

to be due to the so-called RTS/CTS congestion induced

problem. Previous research in TCP and TFRC performance

improvement over wireless networks includes investigating

loss discrimination algorithms (LDA) in order to distinguish

losses due to congestion from those caused by random

wireless errors [19, 20, 21, 22, 23, 24, 25, 26]. Moreover,

several other adaptive RTP-based congestion control schemes

use a similar approach to react to a loss situation in the

network. A first set try to investigate the correlation between

the ROTT (Relative One-way Trip Time) and a congestion

loss. Extensive experimental results conducted in [19, 20]

show that spike-trains observed in a ROTT-graph are only

related to congestion losses and not to random losses.

Congestion control Schemes like PASTRA [27] and VTP [28]

take profit of the ROTT loss discrimination algorithm to find

congestion signals. Another approach, called the inter-arrival

scheme, uses the time between the arrivals of two consecutive

packets as a congestion indication [19, 20]. In [29], Vicente et

al. present the design of LDA+, a loss-delay based congestion

algorithm, based on the inter-arrival scheme. An improvement

of the Datagram Congestion Control Protocol (DCCP) is also

considered in [30] showing that the bandwidth utilization is

improved by more than 30% and up to 50% in significant

setups. The PDA has previously been adapted to ABR flows


congestion control in ATM networks [31, 3

efficiency in terms of fairness, accuracy and

deployed for congestion control for UDP traffic

environments [33]. However, despite their simplicity and transparency, these

implicit flow control approaches can be tricky and unreliable over wireless networks, because large delay fluctuations are inherent to such types of networks. Moreover, some of the previously-cited works report inaccuracies in these differentiators [34].

III. THE PROPOSED CROSS LAYER

The proposed cross layer approach to congestion control in wireless networks can be realized in two stages:

Stage I: A Dual bit Congestion Control Protocol (DBCC) that uses two ECN bits in the IP header of a pair of packets as feedback is used.

Stage II: Secondly, for better QoS and fairshain mobile multimedia wireless networks, a combined mechanism, called the Proportional and Derivative algorithm [PDA] is proposed at the transport layer for UDP traffic congestion control.

This combined approach to dealing with congestion atthe IP and the UDP layers also takes into account only the congestion caused packet losses and helps in mitigating congestion in all wireless networks including encrypted ones very effectively. Each stage is implemented individually and explained below along with simulation results and then unified technique is implemented and the results of the combined technique are discussed in the end.

Figure 1: Architecture of the Initial approach

Figure 2: Architecture of the Cross Layer Approach


32] and proved its

ency in terms of fairness, accuracy and stability when

deployed for congestion control for UDP traffic in wired

However, despite their simplicity and transparency, these can be tricky and unreliable large delay fluctuations are

Moreover, some of the inaccuracies in these

AYER DESIGN

o congestion control in wireless networks can be realized in two stages:

Dual bit Congestion Control Protocol (DBCC) that uses two ECN bits in the IP header of a pair of packets as

Secondly, for better QoS and fairshare of bandwidth mobile multimedia wireless networks, a combined

mechanism, called the Proportional and Derivative algorithm [PDA] is proposed at the transport layer for UDP traffic

This combined approach to dealing with congestion at both the IP and the UDP layers also takes into account only the congestion caused packet losses and helps in mitigating congestion in all wireless networks including encrypted ones

implemented individually and ow along with simulation results and then the

the results of the combined technique are discussed in the end.

Approach

A. The Dual Bit Congestion Control[DBCC]

The design of DBCC is motivated by

First, most feedback based congestion control

require the use of multiple bits in the IP

to headers of the protocols above the IP

deployment challenges in encrypted networks.

Second, most congestion control protocols are designed

wired networks and treat both types of loss as congestion

caused loss. While error-caused losses are typically absent in

wired networks, they are common in wireless networks.

Experiments show that reacting to error

caused loss, can significantly decrease the performance of any

congestion control protocol. Thus, the target operating

environments of DBCC are IP

including encrypted wireless networks.

eight bits of the IP header, two ECN bits and six Type of

Service (ToS) bits, can bypass

are available for end to end signaling.

reserved for signaling differentiated

to congestion control, DBCC will only

of the IP packet header for carrying congestion

signaling feedback.

Overview: Relying on two new schemes,

efficiently in all wireless networks.

1. First and albeit the fact that D

packet four bit representation of the LF, it introduces a packet

ordering management strategy that is quite distinct.

utilizes the information available in the IP header and only

manipulates two existing ECN bits to c

information. The IPID field of the IP header originating from a

host is either monotonically increasing or chosen uniformly at

random. In either case, the LSB of IPID flips over quickly

enough to be used for signaling

DBCC only uses the LSB of the IPID field. Further, the use of

IPID field bits is passive, i.e., the bit values are inspected but

not changed by DBCC. A packet with an LSB value of zero is

used as the MSP and a packet with an LSB value of one is

used as the LSP. If the IPID is increased incrementally, the

LSB bit flips over for any pair of consecutive packets which is

perfect for differentiating MSP

randomly, then DBCC uses the first packet with an LSB value

of zero for carrying MSP and the first packet with an LSB

value of one for carrying LSP. As evidenced in our

experiments, it is safe to assume that bit flips, with a

probability of 0.5, occur quickly enough with respect to

necessary congestion reaction speed especially over larg

networks.

2. Second, DBCC utilizes a heuristic scheme for

differentiating error-caused loss

This heuristic scheme runs at the

maintains the history information of congestion

bottleneck link of a path. Upon detection

scheme makes an identification of the

the saved history information. Given the fact that the feedback

is updated with the receipt of every ACK, it is reasonable to

International Journal of Computer Science and Information Security,


Dual Bit Congestion Control[DBCC]

is motivated by two observations.

First, most feedback based congestion control protocols either

require the use of multiple bits in the IP header or even access

to headers of the protocols above the IP layer, thereby facing

deployment challenges in encrypted networks.

Second, most congestion control protocols are designed for

wired networks and treat both types of loss as congestion

caused losses are typically absent in

tworks, they are common in wireless networks.

show that reacting to error-caused and congestion

loss, can significantly decrease the performance of any

congestion control protocol. Thus, the target operating

are IP-based wireless networks

including encrypted wireless networks. This means that only

two ECN bits and six Type of

Service (ToS) bits, can bypass the encryption boundaries and

are available for end to end signaling. As the ToS bits are

reserved for signaling differentiated services as opposed

will only use the two ECN bits

of the IP packet header for carrying congestion control

Relying on two new schemes, DBCC works

wireless networks.

e fact that DBCC uses a double

of the LF, it introduces a packet

ordering management strategy that is quite distinct. It only

information available in the IP header and only

existing ECN bits to carry congestion

information. The IPID field of the IP header originating from a

is either monotonically increasing or chosen uniformly at

random. In either case, the LSB of IPID flips over quickly

to be used for signaling MSP/LSP. Specifically,

only uses the LSB of the IPID field. Further, the use of

field bits is passive, i.e., the bit values are inspected but

packet with an LSB value of zero is

and a packet with an LSB value of one is

. If the IPID is increased incrementally, the

any pair of consecutive packets which is

MSP from LSP. If it is varied

randomly, then DBCC uses the first packet with an LSB value

and the first packet with an LSB

of one for carrying LSP. As evidenced in our

is safe to assume that bit flips, with a

quickly enough with respect to

speed especially over large BDP

utilizes a heuristic scheme for

caused loss from congestion-caused loss.

This heuristic scheme runs at the transmitting side and

maintains the history information of congestion status over the

link of a path. Upon detection of loss, the heuristic

scheme makes an identification of the source of loss based on

Given the fact that the feedback

the receipt of every ACK, it is reasonable to




assume that the congestion status of a network can be

continuously tracked by the sender. It is especially important

to realize that a congestion caused loss event has a much

longer duration than an error caused loss event. Relying on the

above fact, the heuristic algorithm of DBCC assumes that a

sender can identify the cause of a loss by keeping track of the

status of the network. In order to track the status of the

network, the heuristic algorithm proposes maintaining a

revolving congestion history Bit Map (BM) of size N at the

sending side. Upon the receipt of an ACK, the bit at position

BM(1) is dropped, the bit at position BM(i) with i ∈ 1,…,N is

shifted to the left so it takes the position of bit BM(i − 1), and

the bit at position BM(N) is set to 1 if the new ACK indicates

congestion or otherwise to 0. If at any time, the right most T

consecutive bits with T ≤ N are set to 1 in the bit map, a binary

flag called Congestion Flag (CF) is set to 1. Otherwise, the

flag is set to 0. Upon detection of a loss, if CF flag is set, then

the loss is safely determined as a congestion-caused loss

triggering a multiplicative decrease operation to

cwnd(congestion window). Otherwise, the loss is considered to

be an error-caused loss and the sender simply maintains the

current cwnd. In the case of DBCC, the link LF is

encapsulated in ACK packets and the OVER LOAD represents

a LF beyond 100%. Thus, OVER LOAD is used as the

indicator of congestion.

According to our experiments, setting N to 32 and T to 16

represent optimal choices. We note that with these choices

of values, maintaining the revolving history bit map only

requires 4 bytes of storage on a per flow basis. While N should

essentially be a function of flow cwnd, we set the value of N to

32 for the convenience of implementation. We also note that

the value of cwnd for larger flows could be easily scaled to fit

the 32 bits of N. Fig. 3 illustrates the operation of the heuristic

algorithm of DBCC.

Figure 3: Illustration of the loss differentiating heuristic algorithm in DBCC

3. Finally, the security mode operation IPSec operates in

two modes: transport mode and tunnel mode. In the transport

mode, the original IP header is kept after getting authenticated

by IPSec. Thus, DBCC can still access IPID and ECN bits as

usual in IPSec transport mode. In contrast, the entire packet is

encrypted and authenticated in IPSec tunnel mode. As a result,

the original IP header becomes invisible in the encrypted

packet. Since the LSB bit of the IPID in the original IP header

may not necessarily be the same as the one in the new IP

header, DBCC utilizes the IPID only on the Cipher Text (CT)

side but not on the Plain Text (PT) side for packet ordering.

As DBCC will be installed and configured at the IPSec router,

it is safe to assume that DBCC will have access to both CT

and PT headers of a packet. Specifically, DBCC provides two

router modules: i) Security Module (SM) running only on

IPSec routers that cooperates with IPSec gateways, and ii)

Normal Module (NM) running on both IPSec gateways and

other routers.

Assuming an FTP or a comparable connection has

been established, the flow of events at the IPSec gateways is

as follows:

i) A DBCC packet arrives at the ingress of an IPSec

gateway. Before the packet goes to the IPSec module for

encryption, DBCC SM will first catch the packet, save the

packet ordering information, i.e., MSP/LSP and the value of

the LF as indicated in the ECN bits. Then DBCC SM delivers

the packet to the IPSec module. After the new IP header is

generated and ready to be transmitted through the tunnel,

DBCC SM catches the outgoing packet again and encodes

ECN bits with MSB/LSB bits of the saved LF depending on

the LSB bit of the IPID in the new IP header. Note that, after

the original IP header is encrypted, DBCC has no idea of

whether the new packet is a TCP packet or a packet using

another protocol, e.g., UDP. Thus, DBCC encodes ECN bits

regardless of the original protocol type, which introduces

overhead for non-TCP packets. In fact, this is the tradeoff

between efficiency and protocol complexity. That said, we

note that the resulting overhead is not significant because i) it

is only introduced when transmitting over IPSec tunnels; and

ii) it is only associated with the operations of encoding an LF.

ii) At the output interface of the ingress IPSec gateway,

DBCC NM takes over. DBCC NM compares the LF in the

packet with the LF of its downstream link interface and

updates the LF in the packet if necessary

iii) At the intermediate router on the CT side, DBCC NM

operates as DPCP router module except that DBCC uses the

LSB bit of IPID to identify MSP/LSP.

iv) At the egress of the IPSec gateway and before the

encrypted packet goes to the IPSec module for decryption,

DBCC SM will catch the packet and save the LF value as

indicated by the ECN bits of the packet. Note that after the

packet is decrypted, the IPSec module will copy the ECN bits

from the new IP header to the original IP header on the PT

side. However, the packet ordering information cannot be

simply transferred to the PT side. While DBCC SM can access

both CT and PT side, DBCC SM dedicates to change the

contents of the packet as minimally as possible. Simply put,

DBCC SM does not directly pass any bits from the CT side to

the PT side. Note that, the LSB bit of the IPID in the original

IP header is not necessarily the same as the one in the new IP

header. Thus, instead of changing the value of the LSB bit of

the IPID field in the original IP header for the purpose of

matching the one in the IP header used by the IPSec tunnel,

DBCC uses the relative order of the TCP seq and ack numbers

as the indication of MSP/LSP after the original IP header is

retrieved. In this way, DBCC will not change any bits in the IP

header of the decrypted packet. Furthermore, DBCC SM has

to keep a copy of the LF of the upstream link of the egress




IPSec gateway for each IPSec tunnel. DBCC SM inspects the

ECN bits in the packet and compares it with the MSP/LSP of

the saved copy of the LF of its upstream link. Based on the

results of the comparison, DBCC SM manipulates the seq and

ack numbers in order to mark the packet as MSP or LSP. Then

the packet is delivered to DBCC NM. DBCC NM updates the

ECN bits according to the LF of its downstream link following

the operating mechanism of DPCP.

B. THE PDA CONTROLLER FOR UDP TRAFFIC

This controller helps achieve better QoS and fairshare of

bandwidth in mobile multimedia wireless networks. This PDA

uses quite a standard approach: the level of network

congestion is monitored through the occupancy x of the buffer

which is maintained with the control target being xo. Based on

the difference between x and xo, the PDA controller associated

with each link computes periodically at time n a fair rate q(n)

and forwards the result to the next gateways till the

destination. Using a feedback scheme, a destination supplies

the source with the minimal received q(n). In their turn,

sources have to adapt their transmission rate according to

received fair rate.

1) Buffer equations

Shown in Fig. 4, each node has a congestion controller

associated to its outgoing link i, this controller calculates at

each control period n a supported fair rate qi(n) based on local

information: the difference between the buffer occupancy xi(n)

and a fixed threshold xo, as well as the control decision at

present and in the finite past: qi(n-1), qi(n-2),…… qi(n-k)

Then, the dynamics of buffer i is described by the following equation:

+ 1 = − ∝

− −

, ! ∈ " 1

Where j and k are non negative integers.

Figure 4. A PDA controller.

The saturation function is such that:

#$ = % 0 !' $ < 0 !' $ > $ *ℎ,-.!/,0 The saturation function is introduced to impose bounds on

the computed qi(n): the lower bound zero keeps qi(n) positive,

whereas the upper bound qO limits the sending rate of

connections with non-congested paths.

As stipulated in [19], in order to ensure the system stability,

the coefficients αj and βk must satisfy the following conditions:

∑ 2 > 0, ∑ = 03 2

The first order derivative PDA controller is for the case of j=0, k=0, and is so governed by the following equation: + 1 = 5 −∝ − − 6 3

And, according to conditions stipulated by equation (2), equation (3) leads to: + 1 − 5 −∝ − 6 4

With the above design, the system does not match the

expected stability criteria: first, it exhibits an instable behavior

with several burst losses .Second, the buffer occupancy x does

not oscillate in the neighboring of the threshold xu. These are

confirmed by experimental results, which now demonstrate

the need to introduce a second derivative component in the

PRDR controller equation as shown below:

+ 1 = + + 1 − 9 5

+ 1 = / ;5 −∝ − − 2< − = − 6 6

where ∝ and 2<, are the two first derivative control gains and 9 denotes the rate at which new connections have been

admitted to the network during the time interval [n, n+1].

Solving the equations for proper stability (i.e poles of the

polynomial have negative real roots), the control gains have to

obey the following conditions: −2 < 2< < 0

? ∝+∝<> 0

? ∝<∝<+ 4

The control gains ∝ and ∝< , values can be selected among

the set of values defined in the domain D presented in

Fig.5.An appropriate choice of ∝ and ∝< , will be obtained by

simulations in the next section.




Figure 5. Stability domain of the control gains ∝ and ∝<

2) Adaptation of PDA for UDP flow control

In the transport layer, the Real-Time Protocol (RTP) is

used since it can be implemented on top of UDP/IP stack. For

a feedback mechanism, we chose to use the RTCP protocol,

since it doesn’t consume an additive bandwidth and doesn't

inject too much supplementary traffic in the network. Indeed,

for a sender transmission rate of 2Mbps, RTCP sender reports

sent every 0.1 seconds add only a supplementary traffic of

2.5%. The design of the PDA algorithm is as follows: an

incoming traffic source ‘s’ connected to a destination ‘d’

expresses its initial desired rate Rd in a field of an RTCP

"Application-specific" control packet which is forwarded to

next router until it reaches the destination d. Each intermediate

node m on the path from s to d captures the value of Rd carried

in the RTCP control packet and substitutes it to the locally

computed fair share rate qm if smaller and forwards this

information to its neighbour node m+I. Finally, the control

packet reaches the destination d with the smallest value of qm

on the connection path. Then, the destination sends the

received fair share back to source s in an RTCP RR packet,

which replaces its actual transmission rate Rd by the received

qm.. Different sources periodically send RTCP control packets

every TControl. The choice of the TControl value affects sensibly

both the transient response (settling time and initial connection

parameters) and the control overhead due to the computation

and the transmission of the feedback information. Faster

updates periods lead to shorter settling time, more rapid steady

state and smaller buffer overshoot, whereas a smaller TControl

value increases the control overhead.

IV . PERFORMANCE EVALUATION IN WIRELESS ENVIRONMENT

In this section, simulation studies and experimental studies

of DBCC and PDA are presented, first individually and then

combined.

A. DBCC

DBCC is implemented in both NS-2 simulator and Linux

Kernel. Performance of DBCC, DPCP, and VCP are compared

in terms of efficiency and fairness. Since DBCC is proposed

as an extension of DPCP for encrypted wireless networks, our

target environment is characterized by moderate bandwidth (2

− 10Mbps), low delay (200 − 1000ms) lossy links. The

wireless effects are introduced by utilizing the temporally

correlated Gilbert Elliott (GE) model [35].

We now compare the performance of DPCP and VCP over a

four bottleneck parking lot topology as illustrated by Fig. 6(a).

All of the links have a one-way delay of 250ms and a

bandwidth of 4Mbps except L2 that has a bandwidth of

2Mbps. The GE model is applied on a per link basis in order

to introduce an average loss rate 5% for each link. There are

two types of aggregate FTP flows traversing the topology. The

first type is referred to as a Long Flow and represents the

combined traffic of 30 FTP flows traversing all of the links in

the forward left-to-right direction. The second type is referred

as to as a Local Flow. There are four Local Flows each of

which representing 10 FTP flows traversing each individual

link in the forward direction. Except those flows that traverse

link L2 and start after 1000 seconds, all other Local Flows

start at the beginning of the experiments.

Figure 6. An illustration of a) Parking lot b)Dumbbell topologies

Note that if no wireless loss is introduced, DBCC and

DPCP achieve nearly identical performance as they share

same control policy. With the heuristic scheme, DBCC can

significantly improve the performance of DPCP over a lossy

link.

Fig. 7 shows the bandwidth split ratio of VCP, DPCP,

and DBCC respectively. Ideally, during the first 1000 seconds,

both Long and Local Flows are to equally split the bandwidth

of a shared link. Starting from 1000-th second when an extra

Local Flow starts at link #2, the utilization of Long Flows at

Link #0 should drop to 25% while the utilization of Local

Flows should go up to 75%.

Figure. 7. a) A performance comparison of DBCC, DPCP, and VCP over link

#0. 7. b) A performance comparison of DBCC, DPCP, and VCP over link #2.




In Fig.7a, VCP exhibits a biased fairness characteristic

splitting the bandwidth of link #0 with a ratio of 15 to 1.

While DPCP demonstrates a significantly better fairness

characteristic than VCP, it shows inefficiency in terms of the

bandwidth utilization due to the effect of its reaction to loss. In

contrast, DBCC shows both good fairness and efficiency.

At link #2, we expect to see a near 100% bandwidth utilization

for Long Flows during the first 1000 seconds and a split of

50% in the last 1000 seconds between Long and Local Flow

when the Local Flow joins. As illustrated by Fig. 7b, both

DBCC and DPCP show good fairness and responsiveness,

although DBCC outperforms DPCP in terms of bandwidth

utilization. To the contrary, the bandwidth split ratio does not

change even when Local Flows are turned on in the case of

VCP showing that VCP fails to achieve fairness in high BDP

multiple bottleneck topologies serving flows with

heterogeneous RTTs.

1) Experimental Studies

In this subsection, we describe our implementation of

DBCC in the Linux Kernel. The implementation approach

follows that of VCP as described in [36]. Again, we introduce

packet loss using our GE error model implementation in the

Linux Kernel. In this section, we do present our experimental

study conducted over a real testbed comparing the

performance of VCP, DPCP, and DBCC. We present the

results associated with a single bottleneck scenario. We use a

dumbbell topology (Fig. 7.b) with the settings used for

experiments matching those of [36]. Though not shown here,

the performance of DBCC in multi-bottleneck scenarios

follows the same pattern shown in our simulation studies.

Fig. 8 compares the bandwidth utilization of VCP, DPCP,

and DBCC over the single bottleneck link. In our experiments,

a loss rate of up to 30% is introduced. Thus, both DPCP and

VCP fail to open the cwnd efficiently in the absence of the

heuristic scheme, and therefore exhibit a low utilization

characteristic. Note that while DPCP achieves higher

bandwidth utilization than VCP, it demonstrates oscillations

due to its inappropriate reaction to error-caused loss. The

improvement comes from the faster recovery speed of DPCP

in contrast to VCP. In contrast, DBCC can identify the source

of a loss and ignore error-caused loss. In the figure, DBCC can

achieve significantly better bandwidth utilization than both

DPCP and VCP although it shows oscillations due to the

associated retransmissions and timeouts.

It is clearly seen from the results that DBCC overcomes the

limitations of DPCP by using an alternative packet ordering

management scheme. Rather than accessing the TCP header,

DBCC passively inspected the LSB bit of the IPID field in the

IP packet header to identify whether a packet is the MSP or

LSP in a packet pair sequence. Furthermore, DBCC utilized a

heuristic loss identification scheme to differentiate error-type

and appropriately react to loss.

Figure 8. A performance comparison of DBCC, DPCP, and VCP over the

bottleneck link of our experimental dumbbell topology.

We implemented DBCC in both NS-2 and the Linux Kernel.

Through simulation we demonstrated that the fairness and

efficiency characteristics of DBCC are comparable to those of

DPCP in wired networks. We also demonstrated that in high

BDP networks, both DBCC and DPCP significantly

outperform VCP in terms of fairness and efficiency. As the

main differentiating factors, we showed that i) unlike DPCP,

DBCC can operate over IPSec encrypted networks, and ii)

relying on its heuristic loss identification algorithm, DBCC

can significantly outperform DPCP in wireless environments

characterized by tandem loss

B. PDA Network topology and test configurations

In order to study the performances of the PDA algorithm

in a wireless environment, we considered in our simulations a

heterogeneous topology with base station (BS) on the NS-2.31

network simulator that implements the topology described in

Fig. 9.Here, N sources located in the wired side of the network

initiate k CBR/RTP traffic to N wireless destinations. All

sources are connected to the BS via a gateway G with 15Mbps

and 10 ms-delayed links. The channel bandwidth is 11 Mbps

and the payload size of each data packet is 1500 bytes.

All results are given from five times simulations with 300

seconds duration each. The DSDV routing protocol is used.

The BS implements the second-order PDA, depicted in

equation (5) and (6): for a queue length of 200 packets, the

threshold on queue occupancy is set to 180 packets. Since

the maximal good-throughput of 11 Mbps wireless channels is

about 4.5 Mbps, we judged that a value of 4.5 *0.9 is a good

choice for the target rate. The control period T control is set to

100 ms and the control gain (∝ to 0.8). In our simulations,

the following several important performance metrics are

evaluated:

• Good end-to-end throughput -; the amount of data

delivered to the destination for flow ! = ≤ ! ≤ "

• BS buffer occupancy x

• Stability measured as the standard deviation for -, series,

denoted A • Fairness index: we use the Jain's fairness index used

in [37] and defined by:




' = ∑ -BC D" ∑ -DBC

Figure 9. A PDA controller

1) Channel allocation

Initially consider the network described in Fig.9 where the

BS uses a classical drop-tail queuing algorithm. Fig.10 plots

the instantaneous rates of 4 competing RTP flows, and Fig.11

plots the instantaneous BS buffer occupancy, starting at 0, 20s,

30s and 40s respectively.

It is clear from Fig. 10 and 11 that the wireless channel is

unfairly allocated between the four flows since it represents

the bottleneck of the network. By analyzing the NS-2 trace

file, it was found that the major cause of packet drops is buffer

overflow (IFQ). When the drop-tail queuing discipline is used,

traffic gets synchronized, which allows the first and the fourth

flow to monopolize the queue space and consequently get the

maximal channel bandwidth allocation. The mean value of the

fairness index samples is 0.8. Moreover, the BS buffer is

saturated all the time (x = 200 packets). The unfairness

problem is also revealed for the ten competing RTP flows

depicted in Fig. 12. As we can see, eight flows are

overwhelming the bandwidth (getting almost 500 kbps of

bandwidth), whereas the two remaining flows are roughly

discriminated (obtaining only 200 kbps). In Fig. 13, we

replaced the fourth UDP flow by a single TCP connection. We

can notice that the TCP flow is totally starved since it is

unable to send any packets (TCP rate is zero). The fairness

index is only 0.6. In addition, all the flows see frequent burst

losses (occurring at times 160s, 170s and 260s), which leads to

a simultaneous decrease of the present flow rates, confirming

the hypothesis of the global synchronization problem related

to the drop-tail queuing policy.

Figure 10. Instantaneous rates in the presence of 4 competing RTP flows

Figure 11 Instantaneous buffer occupancy in presence of four competing RTP

flows

2) Channel allocation with II Order PDA algorithm

We present here simulation results with four competing RTP

flows, using the same network configuration for different

values of control gains ∝ and ∝< , within the domain of

Figure 6. The fair rate for the four RTP flows is then:

EF = 4.5HIJ/ × 0.94 = 1.102HIJ/

Figure 12. Instantaneous rates in presence of ten competing RTP flows

Figure 13. Instantaneous rates in presence of three competing RTP flows and

a single TCP connection

The Table I below provides the performance criteria

described in Fig. 10 for different values of ∝ and ∝< . For all

the cases, the measured fairness index f is very close to

1(0.998).




TABLE I Standard Deviations for different values of ∝ and ∝<

As an illustration, Figures 14 and 15 plot the instantaneous

allocated rate for the RTP flows respectively for cases 2 an 7.

From the plots, we can conclude that setting ∝ * 2.5 and ∝< * − 0.5 is the best choice since the system has the best

behavior in terms of stability (not oscillatory) and

convergence. Moreover, the buffer occupancy is better

controlled concerning the fixed threshold xo of 180 packets.

Figure 14. Instantaneous RTP rates or ∝ * 2 and ∝< * − 1

Figure 15. Instantaneous RTP rates for ∝ * 2.5 and ∝< * − 0.5

C. Combined Simulation of DBCC and PDA

Finally both DBCC and the PDA were simultaneously

implemented in the wired cum wireless environment as shown

in Fig.9. The BW Utilization/rate factor significantly improves

when both are implemented together even in highly congested

bottlenecks. As demonstrated in Fig.16 the congestion was

completely eliminated 98% of the time along with the

different flows getting fair allocation of BW.

The fairshare allotted to the 4 different flows is calculated

here also and it was observed that bandwidth utilization is

most efficient at certain specific allotted rates and accordingly

the simulations were conducted for an allotted rate of

5.25Mbps, the new rate of 1.246Mbps was allotted to all the

flows.

EF = 5.25HIJ/ × 0.954 = 1.247HIJ/

The buffer occupancy for different types of RTP flows is as

shown. Results show that our mechanism can fairly allocate

wireless bandwidth resource in heterogeneous networks and

converges to a steady state whenever the input traffic

parameters change.

Figure 16. Instantaneous RTP rates and buffer occupancy for ∝ * 2.5 and ∝< * − 0.5

0

200

400

600

800

1000

1200

1400

0 50 100 150 200 250 300

RTP-1

RTP-2

RTP-3

RTP-4

0

50

100

150

200

250

0 50 100 150 200 250 300

Time (s)

R

a

t

e

k

b

p

s

Buffer

status

Time (s)




V. CONCLUSIONS

Mobile wireless networks present a big challenge to

congestion and flow control algorithms as these links are in a

constant competition to access the shared radio medium. In

this research, a twofold approach using combined Dual bit

Congestion Control [DBCC] at the IP layer and Proportional

and derivative algorithm [PDA] at the Transport layer is used

for more efficient Congestion Control. First, DBCC involving

two ECN bits in the IP header of a pair of packets is used for

congestion situation feedback. This approach differentiates

between the error and congestion-caused losses, and is

therefore capable of operating in all wireless environments

including encrypted wireless networks. Secondly, for better

QoS and fairshare of bandwidth in mobile multimedia

networks, the PDA mechanism is proposed at the transport

layer for UDP traffic congestion control.

Simulation results have shown the efficiency of both

techniques individually in comparison with other standard

existing techniques and also that of the combined technique

where both are implemented together. There is a rate

improvement of about 5.7% as compared to the previous

implementation of the individual techniques and Congestion is

avoided 98% of the time.

REFERENCES

1. Manthos Kazantzidis, Mario Gerla, "End-to-end versus Explicit

Feedback Measurement in 802.11 Networks", in Proceedings of the

Seventh International Symposium on Computers and Communications,

Italy, July 2002.

2. Kai Chen, Klara Nahrstedt, Nitin Vaidya, "The Utility of Explicit

RateBased Flow Control in Mobile Ad Hoc Networks", in Proceeding of

IEEE Wireless Communications and Networking Conference, Atlanta,

USA, March 2004.

3. L. Xu, K. Harfoush, and I. Rhee, “Binary Increase Congestion Control

(BIC) for Fast Long-Distance Networks,” in Proc. of the IEEE

INFOCOM, 2004.

4. I. Rhee and L. Xu, “CUBIC: A New TCP-Friendly High-Speed TCP

Variant,” in Proc. of the PFLDNet’05, Feb. 2005.

5. S. Floyd, “High Speed TCP for Large Congestion Windows,” Aug.

2002.

6. D. Leith and R. Shorten, “H-TCP: TCP for High-speed and Long-

distance Networks,” in Proc. of the PFLDNet’04, Feb. 2004.

7. T. Kelly, “Scalable TCP: Improving Performance in HighSpeed Wide

Area Networks,” Feb. 2003, available at

http://wwwlce.eng.cam.ac.uk/ctk21/scalable/.

8. C. Jin, D. Wei, and S. Low, “FAST TCP: Motivation, Architecture,

Algorithms, Performance,” in Proc. of IEEE INFOCOM, 2004.

9. S. Bhandarkar, S. Jain, and A. Reddy, “Improving TCP Performance in

High Bandwidth High RTT Links Using Layered Congestion Control,”

in Proc. of the FLDNet’05, Feb. 2005.

10. M. Goutelle, Y. Gu, and E. He, “A Survey of Transport Protocols other

than Standard TCP,” 2004,

https://forge.gridforum.org/forum/forum.php?forum id=410.

11. D. Katabi, M. Handley, and C. Rohrs, “Congestion Control for High

Bandwidth-Delay Product Networks,” in Proc. ACM SIGCOMM,

Aug.2002.

12. Y. Xia, L. Subramanian, I. Stoica, and S. Kalyanaraman, “One More Bit

Is Enough,” in Proc. ACM SIGCOMM, 2005, Aug. 2005

13. I. A. Qazi and T. Znati, “On the design of load factor based congestion

control protocols for next-generation networks,” in Proc. of the IEEE

INFOCOM 2008, Apr. 2008.

14. N. Vasic, S. Kuntimaddi, and D. Kostic, “One Bit Is Enough: a

Framework for Deploying Explicit Feedback Congestion Control

Protocols,” in Proc. of the First International Conference on

COMmunication Systems

and NETworkS (COMSNETS), Jan. 2009.

15. X. Li and H. Yousefi’zadeh, “Distributed ECN-Based Congestion

Control,” in Proc. of the IEEE ICC 2009, June 2009.

16. Bansal, H. Balakrishnan, S. Floyd, and S. Shenker, "Dynamic behavior

of slowly-responsive congestion control algorithms," in Proceedings of

ACM SIGCOMM, San Diego, California, USA, August 2001.

17. Y. Yang, M. Kim, and S. S. Lam, "Transient behaviors of tcp friendly

congestion control protocols," in Proceedings of IEEE INFOCOM,

Anchorage, Alaska, USA, April 2001.

18. M. Li, E. Agu, M. Claypool , R. Kinicki, "Performance Enhancement of

TFRC in Wireless Networks", Worcester Polytechnic Institute Technical

Report, December 2003.

19. S. Cen, P. Cosman, , G. Voelker, "End-to-end Differentiation of

Congestion and Wireless Losses", in IEEE/ACM Transactions on

Networking, vol. II, issue 5,2003.

20. C. Parsa and 1. Garcia-Luna-Aceves, "Differentiating congestion vs.

random loss: A method for improving TCP performance over wireless

links" in IEEE WCNC, Chicago, September 2000.

21. J. Liu, I. Matta, and M. Crovella, "End-to-end inference of loss nature in

a hybrid wired/wireless environment," in Proceedings of WiOpt, France,

March 2003.

22. Pierre Geurts, Ibtissam EI Khayat, Guy Leduc, "A machine learning

approach to improve congestion control over wireless computer

networks", In Proceedings of IEEE International Conference on Data

Mining, Brighton, UK, November 2004.

23. Bin Zhou, Cheng Peng Fu Li, V.O.K. , "TFRC Veno: An Enhancement

of TCP Friendly Rate Control over Wired/Wireless Networks", in the

Proceedings of the IEEE International Conference on Network

Protocols, Bejing, China, October 2007.

24. Bin Zhou Cheng Peng Fu Chiew Tong Lau Chuan Ileng Foh, "An

Enhancement of TFRC over Wireless Networks", in Proceedings of the

Wireless Communications and Networking Conference, Hong-Kong,

March 2007.

25. S. Pack, X. Shen, 1. W. Mark, and L. Cai, "A Two-Phase Loss

Differentiation Algorithm for Improving TFRC Performance in IEEE

802.11 WLANs," in IEEE Transactions on Wireless Communications,

February 2007.

26. Neng-Chung Wang , Jong-Shin Chen , Yung-Fa Huang , Chi-Lun

Chiou, "Performance Enhancement of TCP in Dynamic Bandwidth

Wired and Wireless Networks", in Wireless Personal Communications:

An International Journal archive, volume 47, issue 3, November 2008.

27. Y. Tobe, Y. Tamura, A. Molano, S. Ghosh, H. Tokuda, "Achieving

Moderate Fairness for UDP flows by Path-Status Classification", in the

Proceedings of the 25th Annual IEEE Conference on Digital Local

Computer Networks, florida, USA, November 2000.

28. G. Yang, M.Gerla, M.Y; Sanadidi, "T. Smooth and efficient Real-Time

Video Transport in the presence of Wireless errors", in ACM

Transactions on Multimedia computing communications and

Applications (TOMCCAP), vol 2, issue 2, May 2006

29. Vicente E. Mujica V., Dorgham Sisalem, Radu Popescu-Zeletin, Adam

Wolisz, TCP-Friendly Congestion Control over Wireless Networks, in

Proceedings of European Wireless, Barcelona, Spain, February 2004.

30. Ijaz Haider Naqvi , Tanguy Perennou , "A DCCP Congestion Control

Mechanism for Wired- cum-Wireless Environments", in the Proceeding

of the IEEE Wireless Communications and Networking Conference,

Hong-Kong, March 2007.

31. I. Lengliz, F. Kamoun, "A rate-based flow control method for ABR

service in ATM networks", in Computer Networks Journal, Vol. 34,

No.1, pp. 129-138, July 2000.

32. I. Lengliz, F. Kamoun, "System Stability with the PD ABR Flow

Control Algorithm", in the Proceedings of the Fifth IEEE Symposium on

Computer and Communications, ISCC, France, July 2000.

33. I. Lengliz, Abir Ben Ali, F. Kamoun, "A novel issue for multimedia

traffic control in the internet", In Technical Program of Soft'Com, Italy,

October 2004.




34. Kamal Deep Singh, David Ros, Laurent Toutain , Cesar Viho,

"Improvement of Multimedia Streaming using Estimation of Wireless

losses", IRISA Research report, March 2006.

35. X. Li and H. Yousefi’zadeh, “An Implementation and Experimental

Study of the Variable-Structure Congestion Control Protocol (VCP),” in

Proc. Of the IEEE MILCOM, 2007, Oct. 2007.

36. H. Yousefi’zadeh, X. Li, and A. Habibi, “An End-to-End Cross-Layer

Profiling Study of Congestion Control in High BDP Wireless

Networks,” in Proc. of the IEEE WCNC, 2007, Mar. 2007.

37. Ling-Jyh Chen; Chih-Wei Sung; Hao-Hsiang Hung; Sun, T. Cheng-Fu

Chou, "TSProbe: A Link Capacity Estimation Tool for Time-Slotted

Wireless Networks", in Proceedings of IFIP International Conference on

Wireless and Optical Communications Networks, Singapore, July 2007.

Author’s Profiles

Uma Satyanarayana Visweswaraiya is currently pursuing her research

on “Congestion Control Techniques in Communication Networks”

under Dr. K S Gurumurthy in Bangalore University. She is also working

as an Associate Professor in the Eectronics and Communication

Department of RNSIT, Bangalore since 2006, teaching both UG and PG

students in core subjects like Analog and Digital electronics, A & D

Communication, Computer Communication Networks and CMOSRF

Circuit Design. She has published two books, the most recent one being

“Constraint Based design of Communication Networks using GA”,

Lambert Academic Publishing, Germany, 2012, and three papers in

International Journals like Springer-Verlag and IJSER. Her research

interests include Communication Networks and Signal Processing.

Gurumurthy Satyanarayana Rao Kargal has completed his B.E degree in

E & CE, from MCE, HASSAN, Mysore University, and M.E degree

from IIT, ROORKEE and a PhD from IISc, Bangalore-12.

His experience is as Administrator/Coordinator/Specialist and Professor

in ECE at UVCE, BU, Bangalore, INDIA .He was also heading the

department. In addition to teaching and guiding PhD/ME/BE students he

was responsible for the smooth running of the department. He has over

40 international publications to his credit. Presently he is a Professor in

the ECE department of Reva Institute of Technology and Management.

His specialization is VLSI Design and Communication Networks.



The Development of Educational Quality

Administration: a Case of Technical College in

Southern Thailand

Bangsuk Jantawan

Department of Tropical Agriculture and International

Cooperation

National Pingtung University of Science and Technology

Pingtung, Taiwan

Cheng-Fa Tsai

Department of Management Information Systems

National Pingtung University of Science and Technology

Pingtung, Taiwan

Abstract— The purpose of this research were: to survey the

needs of using the information system for educational quality

administration; to develop Information System for Educational

quality Administration (ISEs) in accordance with quality

assessment standard; to study the qualification of ISEs; and to

study satisfaction level of ISEs user. Subsequently, the tools of

study have been employed that there were the collection of 47

questionnaires and 5 interviews to specialist by responsible

officers for Information center of Technical colleges and

Vocational colleges in Southern Thailand. The analysis of

quantitative data has employed descriptive statistics using

mean and standard deviation as the tool of measurement.

Hence, the result was found that most users required software

to search information rapidly (82.89%), software for collecting

data (80.85%) and required Information system which could

print document rapidly and ready for use (78.72%). The ISEs

was created and developed by using Microsoft Access 2007 and

Visual Basic. The ISEs was at good level with the average of

4.49 and SD at 0.5. Users’ satisfaction of this software was at good level with the average of 4.36 and SD at 0.58.

Keywords- Educational Quality Assurance; Educational

Quality Administration; Information System;

I. INTRODUCTION

A. Background

According to the National Education Act (1999) and Vocational Education Act, 2008, the educational institution in Thailand had been changed in various aspect and also called for education reform. Education reform has been caused by social currents of globalization and knowledge. Not only Thailand, but also other countries around the world have focused on the teaching and learning process to students for holistic development. The concept of learning has made personnel of cultural, social, economic and technology for their continuous and timely global trends. The process of teaching and learning quality is important to make a difference. This process must be continuous and consistent with the concept of quality assurance to build up the confidence of students, parents, community and Thai society. Therefore, school administrators along with the teaching methods should confidently make standard ability

of students and the impact on the development of Thailand [1], [2].

The section 47 in the National Education Act (1999) requires the development of quality assurance standards for all educational levels. It includes internal quality assurance and external quality assurance. In the section 48 requires the internal quality assurance as part of a process of the educational institutions continuously. In addition, the section 49 also requires that all educational institutions must obtain external assessment at least once in every five years since the last assessment. The assessment outcomes will be duly submitted to the concerned agencies and the public. Then, educational institutions must prepare to support implementation of the various sections [1],

Hence, researcher interests in development of information management system for administration of educational quality by collecting the data that related with the quality assurance standards. The Technical College and Vocational College in Southern Thailand have to improve a quality of information system and need to collect the data consistently in order to make the decision-maker efficiently. It can identify the weaknesses or problems effectively as well. The remedial measures are needed so as to facilitate subsequent planning and actions required to achieve the goals effectively [3].

B. Research Objectives

The overall objective of this study was to obtain information on the development of educational quality administration for the technical college in southern Thailand. There were four specific objectives of this study following: (1) to survey the needs of using information system for educational quality administration of colleges in southern Thailand; (2) to improve information system for educational quality administration of technical colleges in southern Thailand; (3) to study the quality of information system for educational quality administration of technical colleges in southern Thailand; and (4) to study satisfaction level of information system for educational quality administration of technical colleges in southern Thailand. Actions required achieving the goals effectively [3].



C. Research Hypotheses

This study was based on the hypotheses, which consists of the following three parts: (1) ISEs has appropriated in educational quality management of technical colleges in southern Thailand and good level; (2) ISEs has been developed and has a good quality; and (3) the users of ISEs have a satisfaction in the information system for educational quality management at a high level.

D. Research Scope

As the population, this study was created and develops ISEs of technical colleges in southern Thailand of 47 institutions. Purposive samplings composed of 47 responsible officers for information center of technical colleges and industrial and community colleges in southern Thailand.

As the specialist, there were 3 specialists of quality assessment information system for educational quality management. The details below were the specialist qualifications: (1) the graduate master's degree or higher than that of related fields; (2) the bachelor’s degree or related work in the educational standards of technical colleges and vocational college; and (3) the working-related information and expertise in computer data base of not less than five years, including both sides of eight persons.

E. Research Tools

The tools of study enclosed of four parts: the first tool was the requirement questionnaire of ISEs software. The questionnaire divided was five episodes: (Episode 1) the general data of respondents were multiple choice questions; (Episode 2) the condition used information system in college was multiple choice questions; (Episode 3) the problem of used information system in college was multiple choice questions; (Episode 4) the need of used information system in college was multiple choice questions, and (Episode 5) the opinion and suggestion about needed in used information system program in college; the second tool was the ISEs software; the third tool was the quality evaluate of ISEs; and the fourth tool was the satisfaction questionnaire in used ISEs. It was divided were two parts: (Episode 1) the general data of respondents were multiple choice questions, and (Episode 2) the opinion about used information system was multiple choice questions.

F. Definition Terminology

There were 3 key terminologies of study. Firstly, the technical college in southern regional office of vocational commission: means the institutions technical college and vocational college kind within the technical college in southern regional office of vocational commission in the area south of 47 colleges. Secondly, the user information system: means the personal responsibility information center of technical college in southern regional office of vocational commission. Thirdly, the information system development: means the process of developing software with Microsoft Access and Visual Basic programming a standby Standalone

to be able to store data, Process data and report data According to the information system.

II. METHODOLOGY

A. Data Collection

The process in this study consisted of the information following: (1) the data from questionnaires and the needs of using information system for educational quality administration of colleges in southern regional office of vocational education commission composed of 47 responsible officers; (2) the data from specialists of quality assurance standards, internal quality assurance, and overview information systems composed of eight specialists; and (3) the data from the technical college and vocational college in southern regional office of vocational commission composed of 47 institutions.

B. Data Analysis

1) Create questionnaire needed in used information

system for education quality administration; First step, there were 4 parts: (1) study the principle to

create questionnaire for book document journal related

research, (2) design and create survey of problem and needed to divide issues, (3) take the needed survey of used

information system for education quality administration

draft sent to chief advisor and chief advisor check it and

Edit for suitable, (3) improve and edit needed survey used of

information system for educational quality management of

technical colleges in southern regional office of vocational

education commission, and (4) take need survey in

information system for education quality administration to

use.

2) Development for education quality administration

development information system for education quality

administration to create by procedure; Second step, there were 6 parts: (1) preliminary

investigation, (2) systems analysis, (3) evaluate the consistency of the standard data items, (4) file and database

design (file and database design, program structure design,

and input and output design), (5) system development

(programming, and documentation), and (6) system

implementation.

3) Creating evaluate for education quality

administration have the procedure; Third step, there were 6 parts: (1) study the document ad

related research to education quality administration, (2)

research analysis form objective of research for approach in

used evaluate for comprehensive objective of research, (3)

create evaluate quality of system was evaluation scale 5

levels Likert scale for comments of the specialist of check

system standard and tests information system before to use, (4) take standard evaluate sent to chiefs advisor to check and

suggestion for suitable content, (5) improve and edit

standard evaluate both side according suggestion, and (6)

take quality evaluate of system to sent to the specialist 3

person quality evaluate.



4) Satisfaction questionnaire of information system for

education quality administration cerate to steps; Fourth step, there also were 6 parts: (1) study form book,

document, journal and related research, (2) analysis research

form aims of research for approach in create question to

comprehensive of research, (3) design and create

satisfaction questionnaire according analysis divided 2 parts

such as part 1 general data of respondents , part 2

satisfactions of user in system, (4) take satisfactions

evaluate draft sent to chief advisor and check it and edit for

suitable content, (5) improve and edit evaluate satisfactions according suggestion, and (6) take satisfactions evaluate of

user in information system to give personal user to evaluate.

III. RESULTS

The demand survey of the information system for educational quality management of technical colleges in southern regional office of vocational education commission found that the majority of the information system users demand on quick access to information as a primary demand for 48.94 percent. The secondary demand is the quick search for information for 23.40 percent. Follow by the information arrangement in order for 14.89 percent, aggregate information for 6.38 percent, reduction of information redundancy for 4.26 percent, and reduction of information errors for 2.13 percent, respectively. The majority of the users demand on software that can quickly search the information related to their practices for 82.98 percent, the demand on information recording for 80.85 percent, the demand on software that support their reporting for 17.02 percent, and the demand on software that can reduce the information errors for 2.13, respectively. The users demand on the information system that is capable to print reports efficiently and quickly as a primary demand for 78.72 percent. The secondary demand is the report, produced by the information system, related to their practices for 14.89 percent. The format and details in each report must be easy to interpret for 6.38 percent.

The research results of the information system development for educational quality management of technical colleges in southern regional office of vocational education commission are divided into two parts. Part 1 is the summary of assessment analysis of the correlation between standard and data items with five experts, seven data standards, and 43 indicators. The summary of the validity of 43 questions in the questionnaire survey by five experts found that the correlation value is 1, indicated the correlation of all questions. Part 2 is the result of the system development, using data items that correlated with internal quality assessment standard for the information system structure design and development, including login interface, data input, and data display implemented by Visual Basic and Microsoft Access with the system size 11.7 megabytes.

The summary of the information system quality (Table I), educational quality management of technical colleges in southern regional office of vocational education commission by three experts found the good quality of the information system. According to the quality criteria (mean is 1.5 and

standard deviation is 0.58). The quality in all dimensions are good according to the quality criteria, the quality of the data input is excellent (mean is 4.55 and standard deviation is 0.50), the quality of the results or reports is also excellent (mean is 4.5 and standard deviation is 0.58), the quality of the operational processes is good (mean is 4.30 and standard deviation is 0.50), and the quality of the content is also good (mean is 4.07 and standard deviation is 0.47).

The summary of the satisfaction assessment (Table II), information system using for educational quality management of technical colleges in southern regional office of vocational education commission are as follows: the satisfaction in the data input is high (mean is 4.38 and standard deviation is 0.58), the satisfaction in the contents is high (mean is 4.61 and standard deviation is 0.59), the satisfaction in the operational processes is also the high (mean is 4.63 and standard deviation is 0.61), and the highest satisfaction is in the results or reports (mean is 4.64 and standard deviation 0.53).

TABLE I. THE SUMMARY OF QUALITY ASSESSMENT FOR

INFORMATION SYSTEM

The Summary of Quality Assessment

for Information System X S.D.

Quality

Level

Quality of the data input 4.55 0.50 Excellent

Quality of the content 4.07 0.47 Good

Quality of the operational processes 4.30 0.50 Good

Quality of the results or reports 4.50 0.58 Excellent

Average 4.36 0.58 Good

TABLE II. THE SUMMARY OF SATISFACTION ASSESSMENT FOR

INFORMATION SYSTEM

The Summary of Satisfaction

Assessment for Information System X S.D.

Level of

Satisfaction

Satisfaction of the data input 4.38 0.58 High

Satisfaction of the content 4.61 0.59 Very High

Satisfaction in the operational processes 4.63 0.61 Very High

Satisfaction is in the results or reports 4.64 0.53 Very High

Average 4.52 0.52 Very High

IV. CONCLUSIONS

The demands on the information system for educational quality management of technical colleges in southern regional office of vocational education commission found that the users, who have no experience with the information system for educational quality management, cause the disorder and redundancy information after they worked on it. The information system administrators with less experience want the training on the knowledge about information system. The activities on the information system still lack of supporting software, and this situation causes the problem in data recording that is difficult for the operations. The data search is slow and delaying works. For the results or information reporting, there is no information system that it



can print out to quickly and approved for users’ demand. The college that has complete, efficient, and up-to-date information system to serve the demands can improve the quality of that college efficiently. This development is based on the principles, evidences, and facts that can be proved by the scientific analyses and assessments, logics and causality because the information is required in planning and decision-making that leads to the development of concepts and alternative ways of operation [4].

The development of the information system for educational quality management of technical colleges in southern regional office of vocational education commission found that the correlation can be differentiated into seven dimensions as follows: learners and technical graduates, curriculum and study planning, learner development activities, professional services for publics, research & development, leadership and management, and the standard for the internal quality assurance. Generally, data items are correlated with the standard and the indicators in the internal educational quality assessment.

The information system users’ satisfaction assessment for educational quality management of technical colleges in southern regional office of vocational education commission found that the highest satisfaction of the users is in the results or reports from the information system, high satisfaction in operational processes, contents, and data input.

The suggestions in this study consisted of the following items: Firstly, this study should have a management of information system, quality of education available online, not limited to only one computer at any computer to find information, and access data, anytime, anywhere; and Finally, database and information system of the college should be simple and easy to use, and actions required achieving the goals effectively and efficiently. Responsible officers should be maintaining the database and information system to effective and up date.

V. SUGGESTIONS FOR FURTHER RESEARCH

Further work should be study the behavior of information systems personnel to use information system of the educational quality management of technical colleges in southern regional office of vocational education commission. The research with a similar format should be developing databases on information, report information, and software packages to more easily, such as MySQL database program or on-line system.

ACKNOWLEDGMENT

B. Jantawan would like to express thanks Dr. Cheng-Fa Tsai, professor of the Management Information Systems Department, the Department of Tropical Agriculture and International Cooperation, National Pingtung University of Science and Technology in Taiwan for supporting the outstanding scholarship, and highly appreciates to the Technical College in Southern Thailand for giving the information.

REFERENCES

[1] Office of the National Education Commission. (n.d.). National Education Act of B.E. 2542 (1999). Retrieved November 26, 2011,

from http://www.onec.go.th/Act/5/english/act27.pdf

[2] Vocational Education Act, 2008, Retrieved November 30, 2011, from http://www.ratchakitcha.soc.go.th/DATA/PDF

/2551/A/043/1.PDF

[3] Pinthong, C., 2010, Internal quality assurance standards for the College, and Community Colleges, Department typography Min Buri

School.

[4] Boonreang, K., 1999, Statistical research1, Print No. 7, Bangkok: P.N, printing.


PERFORMANCE EVALUATION OF DATA

COMPRESSION TECHNIQUES VERSUS

DIFFERENT TYPES OF DATA

Doa'a Saad El-Shora

Faculty of Computers and Informatics

Zagazig University

Zagazig, Egypt

Nabil Aly Lashin

Ehab Rushdy Mohamed

Faculty of Computers and Informatics

Zagazig University

Zagazig, Egypt

Ibrahim Mahmoud El- Henawy Faculty of Computers and Informatics Faculty of Computers and Informatics

Zagazig University Zagazig University

Zagazig, Egypt Zagazig, Egypt

Abstract— Data Compression plays an important role in the age

of information technology. It is now very important a part of

everyday life. Data compression has an important application in

the areas of file storage and distributed systems. Because real

world files usually are quit redundant, compression can often

reduce the file sizes considerably, this in turn reduces the needed

storage size and transfer channel capacity. This paper surveys a

variety of data compression techniques spanning almost fifty

years of research. This work illustrates how the performance of

data compression techniques is varied when applying on different

types of data. In this work the data compression techniques:

Huffman, Adaptive Huffman and arithmetic, LZ77, LZW, LZSS,

LZHUF, LZARI and PPM are tested against different types of

data with different sizes. A framework for evaluation the

performance is constructed and applied to these data

compression techniques.

I. INTRODUCTION

Data compression is the art or the science of representing

information in compact form [1]. This compact form is created

by identifying and using structures that exist in the data. Data

can be characters in text files, numbers that are samples of

speech or image waveforms, or sequences of numbers that are

generated by other processes. There are two major families of

compression techniques when considering the possibility of

reconstructing exactly the original source [1], [4]:

1. Lossless compression techniques.

2. Lossy compression techniques.

Figure 1. Lossless compression techniques

Figure 2. Lossy compression techniques

The development of data compression techniques for a variety

of data can be divided into two phases. The first phase is

usually referred to as modeling. In this phase, try to extract

information about any redundancy that exists in the data and

describe the redundancy in the form of model. The second

phase is called coding, in which the difference between the

data and the model are encoded, generally using a binary

alphabet. Having a good model for the data can be useful in

estimating the entropy of the source and lead to more efficient

compression techniques.

There are several types of models:

1. Physical model.

2. Probability model.

3. Markov model.

Physical Model used when knowing something about the

physics of the data generation process. For example, in

speech-related applications However, the physics of data

generation is simply too complicated for developing a model.

Probability Model is the simplest statistical model for the

source is to assume that each letter that is generated by the

source is independent of every other letter, and each occurs

with the same probability. Markov model is one of the most

popular ways of representing dependence in the data,

particularly useful in text compression, where the probability

of the next letter is heavily influenced by the preceding letters.







II. MEASURE OF PERFORMANCE

A compression technique can be evaluated in a number of

different ways.

Measuring the complexity of the technique.

The memory required to implement the technique.

How fast the technique performs on a given machine.

The amount of compression.

How closely the reconstruction resembles the

original.

In this work the Performance evaluation of data compression

techniques concentrated on the last two criteria.

A very logical way of measuring how will a compression

technique compresses a given set of data is to look at the ratio

of bits required to represent the data before compression to the

number of bits required to represent data after compression.

This ratio is called compression ratio [4].

III. DATA COMPRESSION TECHNIQUES

Compression techniques can be divided into two

fundamental and distinct categories: The first techniques are

called statistical compression techniques, as they are statistical

in nature. The second techniques are called dictionary

techniques, they are currently in wide spread use. This

popularity is more due to the fact that the dictionary

techniques are faster and achieve a greater degree of

compression than the statistical compression techniques [12],

[13]. PPM, or prediction by partial matching, is an adaptive

statistical modeling technique based on blending together

different length context models to predict the next character in

the input sequence [14]. The scheme achieves greater

compression than Ziv-Lempel (LZ) dictionary based methods,

which are more widely used because of their simplicity and

faster execution speeds.

A. Statistical Techniques

Statistical compression techniques use the likelihood of a symbol recurring in order to reduce the number of bits needed to store the symbol.

1) Huffman Technique

A more sophisticated and efficient lossless compression

technique is known as “Huffman Coding”, in which the

characters in a data file are converted to a binary code. These

codes are prefix codes and are optimum for a given models

(set of probabilities). Huffman compression is based on two

observations regarding optimum prefix codes: Symbols that

occur more frequently (have a higher probability of

occurrence) will have shorter codewords than symbols that

occur less frequently. The two symbols that occur least

frequently have the same length. The Huffman technique is

obtained by adding a simple requirement to these two

observations. This requirement is that the codewords

corresponding to the two lowest probability symbols differ

only in the last bit. That is, if γ and δ are the two least probable

symbols in an alphabet, and if the codeword for γ was m 0,

the codeword for δ would be m 1. Here m is a string of 1s

and 0s, and denotes concatenation. [2], [3].

2) Adaptive Huffman Technique Huffman coding requires knowledge of the probabilities of the source sequence. If the knowledge is not available, Huffman coding becomes two –pass procedure: the statistics are collected in the first pass, and the source is encoded in the second pass. In order to convert this technique into a one –pass procedure, techniques for adaptively developing the Huffman code were developed based on the statistics of the symbols already encountered. Theoretically, to encode the (k+1)

th

symbol using the statistics of the first k symbols, it is required to compute the code using Huffman coding procedure each time a symbol is transmitted. However, this would not be a very practical approach due to the large amount of computation involved. Adaptive Huffman coding solved this problem [1].

In the adaptive Huffman coding procedure, neither transmitter nor receiver knows anything about the statistics of the source sequence at the start of the transmission. The tree at both the transmitter and the receiver consists of a single node that corresponds to all symbols not yet transmitted and has a weight of zero. As transmission progresses, nodes corresponding to symbols transmitted will be added to the tree, and the tree is reconfigured using an update procedure. Before the beginning of transmission, a fixed code of each symbol is agreed upon between transmitter and receiver [1], [4].

3) Arithmetic Technique

It is more efficient to generate codewords for groups or

sequences of symbols rather than generating a separate

codewords for each symbol in the sequence. However, this

approach becomes impractical for obtaining Huffman codes

for long sequences of symbols. In order to Huffman codes

particular sequences of length m, this needs making

codewords for all possible sequences of length m. This fact

causes an exponential growth in the size of the codebook. It is

desirable to assign codewords to particular sequences without

having to generate codes for all sequences of that length. The

arithmetic coding technique fulfills these requirements. In

arithmetic coding a unique identifier or tag is generated for the

sequence to be encoded. This tag corresponds to a binary

fraction, which becomes the binary code for the sequence [3],

[4].

B. Dictionary Techniques

In many applications, the output of the source consists of

recurring patterns. A classic example is a text source in which

certain patterns or words recur currently. Also, there are

certain patterns that do not occur or with great rarity

occurring. A very reasonable approach to encode these sources

is to keep a list, or dictionary, of frequently occurring patterns.

When these patterns appear in the source output, they are

encoded with reference to the dictionary. If the patterns do not

appear in the dictionary, then it can be encoded using other,

less efficient method. In effect, the input is divided into two

classes, frequently occurring patterns and infrequently

occurring patterns [9], [10].



1) LZ77 Technique

Lempel-Ziv [1977], or LZ77 is an adaptive dictionary-

based compression techniques. LZ77 exploits the fact that

words and phrases within a text file are likely to be repeated.

When there is repetition, they can be encoded as a pointer to

an earlier occurrence, with the pointer accompanied by the

number of characters to be matched. The encoder examines

the input sequence through a sliding window. This window

consists of two parts a search buffer that contains a portion of

the recently encoded sequence, and a look-ahead buffer that

contains the next portion of the sequence to be encoded. In

practice the sizes of the two buffers are larger [15]. The code

encoded as a triple (o, l, c) where o is the offset (The distance

of the pointer from the look-ahead buffer), l is the length of

the longest match and c is the codeword corresponding to the

symbol in the look-ahead buffer that follows the match. It is a

very simple adaptive scheme that requires no prior knowledge

of the source and seems to require no assumptions about the

characteristics of the source [3], [4].

2) LZW Technique

LZW is a universal lossless data compression

technique created by Abraham Lempel, Jacob Ziv , and

Terry Welch[16], [17]. This technique is simple to be

implemented, and has the potential for very high throughput in

hardware implementations [6]. LZW compression creates a

table of strings commonly occurring in the data being

compressed, and replaces the actual data with references into

the table. The table is formed during compression at the same

time at which the data is encoded and during decompression at

the same time as the data is decoded [9]. LZW is a technique

for removing the necessity of encoding the second element of

the pair (i, c). That is, the encoder would only send the index

to the dictionary. So the dictionary has to be primed with all

the letters of the source alphabet. The technique is surprisingly

simple; it replaces strings of characters with single codes. It

does not do any analysis of the incoming text [5].

3) LZSS Technique

This scheme is initiated by Ziv and Lempel [18], [19]. An

implementation using a binary tree is proposed by Bell. The

technique is quite simple: A ring buffer is kept, which initially

contains “space” characters only. Several letters are read from

the file to the buffer. Then the buffer will be searched for the

longest string that matches letters just read, and its length and

position in the buffer will be sent. If the buffer size is 4096

bytes, the position can be encoded in 12 bits. If the match

length is represented in four bits, the <position, length> pair is

two bytes long. If the longest match is no more than two

characters, then just one character is sent without encoding,

and the process is restarted with the next letter.

One extra bit must be sent each time to tell the decoder

whether a <position, length> pair is sent or the code of the

character [4].

4) LZARI Technique

In each step the LZSS technique sends either a character

or a [position, length] pair. Among these, perhaps character

“e” appears more frequently than “x”, and a [position, length]

pair of length 3 might be commoner than one of length 18.

Thus, if the more frequent will be encoded in fewer bits and

less frequent in more bits, the total length of the encoded text

will be diminished. This compression suggests that it should

use arithmetic coding, preferably of adaptive kind, along with

LZSS [4], [7].

5) LZHUF Technique

LZHUF, the technique of Haruyasu Yoshizaki replaces

LZARI’s adaptive arithmetic coding with adaptive Huffman.

LZHUF encodes the most significant 6 bits of the position in

its 4096-byte buffer by table lookup. More recent, and hence

more probable, positions are coded in fewer bits. On the other

hand, the remaining 6 bits are sent verbatim. Because

Huffman coding encodes each letter into a fixed number of

bits, table lookup can be easily implemented [7].

C. PPM Techniques

PPM, or prediction by partial matching, is an adaptive

statistical modeling technique based on blending together

different length context models to predict the next character in

the input sequence. A Series of improvements was described

called PPMC that is tuned to improve compression and

increase execution speed. Also the use of exclusion principle

is used to improve the performance. PPM relies on arithmetic

coding to obtain very good compression performance. PPM is

a combination of several fixed-order context models to predict

the next character in an input sequence. The prediction

probabilities for each context in the model are calculated by

frequency counts, which are updated adaptively and the

symbols that occurs are encoded relative to their predicated

distribution using arithmetic coding [10].

1) PPMC Technique

PPMC (prediction by partial matching without exclusion)

is a technique to assign probability to the escape character is

called the technique C and will be as follows: at any level,

with the current context, let the total number of symbols seen

previous be nt and let nd be the total number of distinct

contexts. Then the probability of the escape character is given

by nd/ ( nd+nt). Any character which appeared in this context nc

times will have a probability nc/( nd+nt).

The intuitive explanation of this technique, based on

experimental evidence, is that if many distinct contexts are

encountered, then the escape character will have higher

probability but if these distinct contexts tend to appear too

many times, then the probability of the escape character

decreases. The PPM technique using technique C for

probability estimation is called PPMC technique.

2) PPMC with Exclusion Technique

PPMC can be modified by using exclusion, this

modification will improve the compression ratio but it is

slower than the first type. Exclusion principle states that: If a



context at a lower level is a suffix of a context at a higher

level, this context is excluded at the lower level [11].

IV. THE EXPERIMENTAL RESULTS AND DISCUSSION

Data compression techniques: Statistical, dictionary and

PPMC are applied on different sizes of standard files, text file,

document file, PowerPoint file and Portable Document Format

file. These experimental tests are carried out to show how the

performance of the technique is varied when dealing with

these file formats. The reduction in size is used to evaluate the

performance.

A. Data compression techniques versus standard files

Corpus is used to evaluate the practical performance of

various text compression schemes. Eight different types of text

are represented, and to confirm that the performance of

schemes is consistent for any given type, many of the types

have more than one representative [2, 8]. The results of

applying the data compression techniques on standard

files are presented in the Table I, Table II, and Table III.

Fig.3 summarizes the results.

TABLE I. RESULTS OF APPLYING STATISICAL TECHNIQUES

AGAINST CORPUS.

Table I, Table II, Table III and Fig. 3 illustrate that all data

compression techniques achieve a reasonable and close

results against standard files. The family of PPM gives

outperforms over others data compression techniques and The

Huffman technique is the least performer with this file

format.

TABLE II. RESULTS OF APPLYING DICTIONARY TECHNIQUES

AGAINST CORPUS.

TABLE III. RESULTS OF RUNNING PPM TECHNIQUES AGAINST CORPUS.

Figure 3. Results of applying data compression techniques on standard files

B. Data compression techniques versus document files

Data compression techniques are tested against ten document

files with different sizes. The result is illustrated in Fig. 4.



Figure 4. Results of applying data compression techniques on document files

From Fig. 4, it is founded that the performance of statistical

techniques is convergent, and the arithmetic technique

achieves a little bit better results. The family of PPM

especially PPMC with exclusion achieves the best results

compared to other techniques, while the worst one is LZW.

A slight improvement in performance is achieved with

LZARI, and LZHUF compared to other dictionary techniques.

C. Data compression techniques versus text files

Data compression techniques are tested against selected ten

text files with different sizes. The result is illustrated in Fig. 5.

Figure 5. Results of applying data compression techniques on text files

From Fig. 5, it is pointed out that the statistical techniques

achieve very convergent results, while the results achieved

with LZARI is better than that of other dictionary techniques

when dealing with text files. Also, PPM family is proved to be

the best when dealing with this type of data. The performance

of PPMC with exclusion is a little bit better compared to

PPMC without exclusion.

D. Data compression techniques versus powerpoint files

Ten PowerPoint files with different sizes are selected and used

to evaluate the performance of the data compression

techniques. The result of running the data compression

techniques against the selected PowerPoint files is shown in

Fig.6.

Figure 6. Results of applying data compression techniques on PowerPoint

files

From Fig. 6, it is very obvious that all techniques except

PPMC with exclusion ascertain very bad results when dealing

with PowerPoint files, LZW gives negative result, as it does

expansion rather than compression. On the contrary, the

performance of PPMC with exclusion is improved.

E. Data compression techniques versus portable document

format files

The data compression techniques are tested on collected ten

portable document format files with different sizes and the

result is elucidated in Fig. 7.

Figure 7. Results of applying data compression techniques on portable

document format files

From Fig. 7, it is accentuated that the performance of all data

compression techniques except PPMC with exclusion worsens

dramatically when dealing with portable document files, as the

reduction in size achieved is considerably low. Negative

results are founded with LZW and LZSS, as they do expansion

in file size rather than compression.

The overall performance of applying data compression

techniques against all files format are concluded in Fig. 8.

Fig. 8 states that best results are achieved with using PPM

family especially PPMC with exclusion compared to other

techniques. The statistical techniques are proved to be the

worst ones and their performance is convergent.



Figure 8. The overall performance of data compression techniques

versus all files format

Also, the performance of all data compression techniques is

converged when dealing with standard files, document files

and text files. On the other side, an enormous difference in

performance is achieved when dealing with PowerPoint files

and Portable Document Format files, as a noticeable

improvement in performance is achieved with PPMC with

exclusion, while disgraceful results are achieved with other

techniques. Finally, negative results are obtained with both

LZW when dealing with PowerPoint files and Portable

Document Format files and LZSS when dealing with Portable

Document Format files.

CONCLUSION

In this research, the performance of the data compression

techniques; Huffman, Adaptive Huffman and arithmetic,

LZ77, LZW, LZSS, LZHUF, LZARI, PPMC without

exclusion and PPMC with exclusion is tested and

experimented against standard files, text files, document files,

PowerPoint files and Portable Document Format files are

presented. The amount of compression is selected to be a

logical way of measuring the performance of these techniques.

From the experimental results, it is concluded that:

The performance of data compression techniques is varied

considerably when applying on different types of data.

PPMC with exclusion achieves the best results with all file

formats than other techniques.

Statistical techniques give the least results comparing to the

other techniques. Nevertheless, they open the door to a great

revolution in the data compression field as they are

considered as the backbone of all advanced techniques. All data compression techniques achieve good and

convergent results with standard file, text file and document

file.

When dealing with PowerPoint files and portable document

format files, a very poor performance is achieved with

LZ77, LZW, LZSS, LZHUF, LZARI, and PPMC without

exclusion, by contrast, a great leap in performance is

achieved with PPMC with exclusion.

LZW achieves Negative results when applying on Portable

Document Format files and PowerPoint files, also, LZSS

with Portable Document Format files.

It is highly recommended for users to select PPMC with

exclusion when working with Portable Document Format and

PowerPoint files. When dealing with text and document files,

users can select any technique as their performance is

convergent, PPM family achieves a little bit good result than

others.

REFERENCES

[1] Pu, I.M., Fundamental Data Compression, Elsevier, Britain, 2006.

[2] Bell, T.C., Cleary, J.G. and Witten, I.H. Text compression. Prentice Hall, Englewood Cliffs, NJ, 1990.

[3] Senthil Shanmugasundaram, Robert Lourdusamy”A Comparative

Study Of Text Compression Algorithms". International Journal of Wisdom Based Computing, Vol. 1 (3), December 2011.

[4] Khalid Sayood, “Introduction to Data Compression”, 2nd Edition, San

Francisco, CA, Morgan Kaufmann, 2000.

[5] Haroon A, Mohammed A “Data Compression Techniques on Text

Files: A Comparison Study” International Journal of Computer

Applications (0975 – 8887) Volume 26– No.5, July 2011. [6] Draft Lecture Notes "compression Algorithms: Huffman and Lempel-

Ziv-Welch (Lzw)". Last Update: February 13, 2012.

[7] H. Okumura,” Data Compression Algorithms of Larc and LHarc,” GameDev. Net, Yokosuka, Japan, 1999.

[8] Bell, T.C., Witten, I.H. and Cleary, J.G. "Modeling for text

compression," Computing Surveys 21(4): 557-591; December 1989.

[9] Lenat, Doug, Lempel-Ziv compression, http://foldoc.doc.ic.ac.uk/foldoc/foldoc.cgi?Lempel-Ziv+compression

1999.

[10] P. Fenwick,” Block-Sorting Text Compression- Final Report,”

Technical Report 130, Univ. of Auckland, New Zealand, Department of Computer science, 1996.

[11] A. Moffat,” Implementing the PPM Data compression Schemes,” IEEE Trans. Comm., 1990.

[12] K. Sayood,” Introduction to Data Compression,” Morgan

Kaufmann Publishers, 1996.

[13] Klein, D.E.,” Dynamic Huffman coding,” J. Algorithms, 6, pp. 163-180, 1985.

[14] A. Compos,” Finite context modeling,” http:// www. Arturocampos.com, 2000.

[15] A. Compos,” LZ77 the basics of Compression,” http://www.

arturocampos.com, 1999.

[16] M. Nelson,” LZW Data Compression,” Dr. Dobb’s Journal, 1989.

[17] Welch, T.A., “A technique for high-performance data compression,” IEEE computer, 17, no. 6, pp. 8-19, 1984.

[18] H. Okumura,” Data Compression Algorithms of Larc and LHarc,” GameDev. Net, Yokosuka, Japan, 1999.

[19] Ziv and A. Lempel, “A Universal Algorithm for Sequential Data

Compression,” IEEE Transaction on Information Theory, Vol. 23, pp. 337-343, 1977.



(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11 No. 12, December 2013

IJCSIS AUTHORS’ & REVIEWERS’ LIST Assist Prof (Dr.) M. Emre Celebi, Louisiana State University in Shreveport, USA

Dr. Lam Hong Lee, Universiti Tunku Abdul Rahman, Malaysia

Dr. Shimon K. Modi, Director of Research BSPA Labs, Purdue University, USA

Dr. Jianguo Ding, Norwegian University of Science and Technology (NTNU), Norway

Assoc. Prof. N. Jaisankar, VIT University, Vellore,Tamilnadu, India

Dr. Amogh Kavimandan, The Mathworks Inc., USA

Dr. Ramasamy Mariappan, Vinayaka Missions University, India

Dr. Yong Li, School of Electronic and Information Engineering, Beijing Jiaotong University, P.R. China

Assist. Prof. Sugam Sharma, NIET, India / Iowa State University, USA

Dr. Jorge A. Ruiz-Vanoye, Universidad Autónoma del Estado de Morelos, Mexico

Dr. Neeraj Kumar, SMVD University, Katra (J&K), India

Dr Genge Bela, "Petru Maior" University of Targu Mures, Romania

Dr. Junjie Peng, Shanghai University, P. R. China

Dr. Ilhem LENGLIZ, HANA Group - CRISTAL Laboratory, Tunisia

Prof. Dr. Durgesh Kumar Mishra, Acropolis Institute of Technology and Research, Indore, MP, India

Jorge L. Hernández-Ardieta, University Carlos III of Madrid, Spain

Prof. Dr.C.Suresh Gnana Dhas, Anna University, India

Mrs Li Fang, Nanyang Technological University, Singapore

Prof. Pijush Biswas, RCC Institute of Information Technology, India

Dr. Siddhivinayak Kulkarni, University of Ballarat, Ballarat, Victoria, Australia

Dr. A. Arul Lawrence, Royal College of Engineering & Technology, India

Mr. Wongyos Keardsri, Chulalongkorn University, Bangkok, Thailand

Mr. Somesh Kumar Dewangan, CSVTU Bhilai (C.G.)/ Dimat Raipur, India

Mr. Hayder N. Jasem, University Putra Malaysia, Malaysia

Mr. A.V.Senthil Kumar, C. M. S. College of Science and Commerce, India

Mr. R. S. Karthik, C. M. S. College of Science and Commerce, India

Mr. P. Vasant, University Technology Petronas, Malaysia

Mr. Wong Kok Seng, Soongsil University, Seoul, South Korea

Mr. Praveen Ranjan Srivastava, BITS PILANI, India

Mr. Kong Sang Kelvin, Leong, The Hong Kong Polytechnic University, Hong Kong

Mr. Mohd Nazri Ismail, Universiti Kuala Lumpur, Malaysia

Dr. Rami J. Matarneh, Al-isra Private University, Amman, Jordan

Dr Ojesanmi Olusegun Ayodeji, Ajayi Crowther University, Oyo, Nigeria

Dr. Riktesh Srivastava, Skyline University, UAE

Dr. Oras F. Baker, UCSI University - Kuala Lumpur, Malaysia

Dr. Ahmed S. Ghiduk, Faculty of Science, Beni-Suef University, Egypt

and Department of Computer science, Taif University, Saudi Arabia

Mr. Tirthankar Gayen, IIT Kharagpur, India

Ms. Huei-Ru Tseng, National Chiao Tung University, Taiwan


Prof. Ning Xu, Wuhan University of Technology, China

Mr Mohammed Salem Binwahlan, Hadhramout University of Science and Technology, Yemen

& Universiti Teknologi Malaysia, Malaysia.

Dr. Aruna Ranganath, Bhoj Reddy Engineering College for Women, India

Mr. Hafeezullah Amin, Institute of Information Technology, KUST, Kohat, Pakistan

Prof. Syed S. Rizvi, University of Bridgeport, USA

Mr. Shahbaz Pervez Chattha, University of Engineering and Technology Taxila, Pakistan

Dr. Shishir Kumar, Jaypee University of Information Technology, Wakanaghat (HP), India

Mr. Shahid Mumtaz, Portugal Telecommunication, Instituto de Telecomunicações (IT) , Aveiro, Portugal

Mr. Rajesh K Shukla, Corporate Institute of Science & Technology Bhopal M P

Dr. Poonam Garg, Institute of Management Technology, India

Mr. S. Mehta, Inha University, Korea

Mr. Dilip Kumar S.M, University Visvesvaraya College of Engineering (UVCE), Bangalore University,

Bangalore

Prof. Malik Sikander Hayat Khiyal, Fatima Jinnah Women University, Rawalpindi, Pakistan

Dr. Virendra Gomase , Department of Bioinformatics, Padmashree Dr. D.Y. Patil University

Dr. Irraivan Elamvazuthi, University Technology PETRONAS, Malaysia

Mr. Saqib Saeed, University of Siegen, Germany

Mr. Pavan Kumar Gorakavi, IPMA-USA [YC]

Dr. Ahmed Nabih Zaki Rashed, Menoufia University, Egypt

Prof. Shishir K. Shandilya, Rukmani Devi Institute of Science & Technology, India

Mrs.J.Komala Lakshmi, SNR Sons College, Computer Science, India

Mr. Muhammad Sohail, KUST, Pakistan

Dr. Manjaiah D.H, Mangalore University, India

Dr. S Santhosh Baboo, D.G.Vaishnav College, Chennai, India

Prof. Dr. Mokhtar Beldjehem, Sainte-Anne University, Halifax, NS, Canada

Dr. Deepak Laxmi Narasimha, Faculty of Computer Science and Information Technology, University of

Malaya, Malaysia

Prof. Dr. Arunkumar Thangavelu, Vellore Institute Of Technology, India

Mr. M. Azath, Anna University, India

Mr. Md. Rabiul Islam, Rajshahi University of Engineering & Technology (RUET), Bangladesh

Mr. Aos Alaa Zaidan Ansaef, Multimedia University, Malaysia

Dr Suresh Jain, Professor (on leave), Institute of Engineering & Technology, Devi Ahilya University, Indore

(MP) India,

Dr. Mohammed M. Kadhum, Universiti Utara Malaysia

Mr. Hanumanthappa. J. University of Mysore, India

Mr. Syed Ishtiaque Ahmed, Bangladesh University of Engineering and Technology (BUET)

Mr Akinola Solomon Olalekan, University of Ibadan, Ibadan, Nigeria

Mr. Santosh K. Pandey, Department of Information Technology, The Institute of Chartered Accountants of

India

Dr. P. Vasant, Power Control Optimization, Malaysia

Dr. Petr Ivankov, Automatika - S, Russian Federation


Dr. Utkarsh Seetha, Data Infosys Limited, India

Mrs. Priti Maheshwary, Maulana Azad National Institute of Technology, Bhopal

Dr. (Mrs) Padmavathi Ganapathi, Avinashilingam University for Women, Coimbatore

Assist. Prof. A. Neela madheswari, Anna university, India

Prof. Ganesan Ramachandra Rao, PSG College of Arts and Science, India

Mr. Kamanashis Biswas, Daffodil International University, Bangladesh

Dr. Atul Gonsai, Saurashtra University, Gujarat, India

Mr. Angkoon Phinyomark, Prince of Songkla University, Thailand

Mrs. G. Nalini Priya, Anna University, Chennai

Dr. P. Subashini, Avinashilingam University for Women, India

Assoc. Prof. Vijay Kumar Chakka, Dhirubhai Ambani IICT, Gandhinagar ,Gujarat

Mr Jitendra Agrawal, : Rajiv Gandhi Proudyogiki Vishwavidyalaya, Bhopal

Mr. Vishal Goyal, Department of Computer Science, Punjabi University, India

Dr. R. Baskaran, Department of Computer Science and Engineering, Anna University, Chennai

Assist. Prof, Kanwalvir Singh Dhindsa, B.B.S.B.Engg.College, Fatehgarh Sahib (Punjab), India

Dr. Jamal Ahmad Dargham, School of Engineering and Information Technology, Universiti Malaysia Sabah

Mr. Nitin Bhatia, DAV College, India

Dr. Dhavachelvan Ponnurangam, Pondicherry Central University, India

Dr. Mohd Faizal Abdollah, University of Technical Malaysia, Malaysia

Assist. Prof. Sonal Chawla, Panjab University, India

Dr. Abdul Wahid, AKG Engg. College, Ghaziabad, India

Mr. Arash Habibi Lashkari, University of Malaya (UM), Malaysia

Mr. Md. Rajibul Islam, Ibnu Sina Institute, University Technology Malaysia

Professor Dr. Sabu M. Thampi, .B.S Institute of Technology for Women, Kerala University, India

Mr. Noor Muhammed Nayeem, Université Lumière Lyon 2, 69007 Lyon, France

Dr. Himanshu Aggarwal, Department of Computer Engineering, Punjabi University, India

Prof R. Naidoo, Dept of Mathematics/Center for Advanced Computer Modelling, Durban University of

Technology, Durban,South Africa

Prof. Mydhili K Nair, M S Ramaiah Institute of Technology(M.S.R.I.T), Affliliated to Visweswaraiah

Technological University, Bangalore, India

M. Prabu, Adhiyamaan College of Engineering/Anna University, India

Mr. Swakkhar Shatabda, Department of Computer Science and Engineering, United International University,

Bangladesh

Dr. Abdur Rashid Khan, ICIT, Gomal University, Dera Ismail Khan, Pakistan

Mr. H. Abdul Shabeer, I-Nautix Technologies,Chennai, India

Dr. M. Aramudhan, Perunthalaivar Kamarajar Institute of Engineering and Technology, India

Dr. M. P. Thapliyal, Department of Computer Science, HNB Garhwal University (Central University), India

Dr. Shahaboddin Shamshirband, Islamic Azad University, Iran

Mr. Zeashan Hameed Khan, : Université de Grenoble, France

Prof. Anil K Ahlawat, Ajay Kumar Garg Engineering College, Ghaziabad, UP Technical University, Lucknow

Mr. Longe Olumide Babatope, University Of Ibadan, Nigeria

Associate Prof. Raman Maini, University College of Engineering, Punjabi University, India


Dr. Maslin Masrom, University Technology Malaysia, Malaysia

Sudipta Chattopadhyay, Jadavpur University, Kolkata, India

Dr. Dang Tuan NGUYEN, University of Information Technology, Vietnam National University - Ho Chi Minh

City

Dr. Mary Lourde R., BITS-PILANI Dubai , UAE

Dr. Abdul Aziz, University of Central Punjab, Pakistan

Mr. Karan Singh, Gautam Budtha University, India

Mr. Avinash Pokhriyal, Uttar Pradesh Technical University, Lucknow, India

Associate Prof Dr Zuraini Ismail, University Technology Malaysia, Malaysia

Assistant Prof. Yasser M. Alginahi, College of Computer Science and Engineering, Taibah University,

Madinah Munawwarrah, KSA

Mr. Dakshina Ranjan Kisku, West Bengal University of Technology, India

Mr. Raman Kumar, Dr B R Ambedkar National Institute of Technology, Jalandhar, Punjab, India

Associate Prof. Samir B. Patel, Institute of Technology, Nirma University, India

Dr. M.Munir Ahamed Rabbani, B. S. Abdur Rahman University, India

Asst. Prof. Koushik Majumder, West Bengal University of Technology, India

Dr. Alex Pappachen James, Queensland Micro-nanotechnology center, Griffith University, Australia

Assistant Prof. S. Hariharan, B.S. Abdur Rahman University, India

Asst Prof. Jasmine. K. S, R.V.College of Engineering, India

Mr Naushad Ali Mamode Khan, Ministry of Education and Human Resources, Mauritius

Prof. Mahesh Goyani, G H Patel Collge of Engg. & Tech, V.V.N, Anand, Gujarat, India

Dr. Mana Mohammed, University of Tlemcen, Algeria

Prof. Jatinder Singh, Universal Institutiion of Engg. & Tech. CHD, India

Mrs. M. Anandhavalli Gauthaman, Sikkim Manipal Institute of Technology, Majitar, East Sikkim

Dr. Bin Guo, Institute Telecom SudParis, France

Mrs. Maleika Mehr Nigar Mohamed Heenaye-Mamode Khan, University of Mauritius

Prof. Pijush Biswas, RCC Institute of Information Technology, India

Mr. V. Bala Dhandayuthapani, Mekelle University, Ethiopia

Dr. Irfan Syamsuddin, State Polytechnic of Ujung Pandang, Indonesia

Mr. Kavi Kumar Khedo, University of Mauritius, Mauritius

Mr. Ravi Chandiran, Zagro Singapore Pte Ltd. Singapore

Mr. Milindkumar V. Sarode, Jawaharlal Darda Institute of Engineering and Technology, India

Dr. Shamimul Qamar, KSJ Institute of Engineering & Technology, India

Dr. C. Arun, Anna University, India

Assist. Prof. M.N.Birje, Basaveshwar Engineering College, India

Prof. Hamid Reza Naji, Department of Computer Enigneering, Shahid Beheshti University, Tehran, Iran

Assist. Prof. Debasis Giri, Department of Computer Science and Engineering, Haldia Institute of Technology

Subhabrata Barman, Haldia Institute of Technology, West Bengal

Mr. M. I. Lali, COMSATS Institute of Information Technology, Islamabad, Pakistan

Dr. Feroz Khan, Central Institute of Medicinal and Aromatic Plants, Lucknow, India

Mr. R. Nagendran, Institute of Technology, Coimbatore, Tamilnadu, India

Mr. Amnach Khawne, King Mongkut’s Institute of Technology Ladkrabang, Ladkrabang, Bangkok, Thailand


Dr. P. Chakrabarti, Sir Padampat Singhania University, Udaipur, India

Mr. Nafiz Imtiaz Bin Hamid, Islamic University of Technology (IUT), Bangladesh.

Shahab-A. Shamshirband, Islamic Azad University, Chalous, Iran

Prof. B. Priestly Shan, Anna Univeristy, Tamilnadu, India

Venkatramreddy Velma, Dept. of Bioinformatics, University of Mississippi Medical Center, Jackson MS USA

Akshi Kumar, Dept. of Computer Engineering, Delhi Technological University, India

Dr. Umesh Kumar Singh, Vikram University, Ujjain, India

Mr. Serguei A. Mokhov, Concordia University, Canada

Mr. Lai Khin Wee, Universiti Teknologi Malaysia, Malaysia

Dr. Awadhesh Kumar Sharma, Madan Mohan Malviya Engineering College, India

Mr. Syed R. Rizvi, Analytical Services & Materials, Inc., USA

Dr. S. Karthik, SNS Collegeof Technology, India

Mr. Syed Qasim Bukhari, CIMET (Universidad de Granada), Spain

Mr. A.D.Potgantwar, Pune University, India

Dr. Himanshu Aggarwal, Punjabi University, India

Mr. Rajesh Ramachandran, Naipunya Institute of Management and Information Technology, India

Dr. K.L. Shunmuganathan, R.M.K Engg College , Kavaraipettai ,Chennai

Dr. Prasant Kumar Pattnaik, KIST, India.

Dr. Ch. Aswani Kumar, VIT University, India

Mr. Ijaz Ali Shoukat, King Saud University, Riyadh KSA

Mr. Arun Kumar, Sir Padam Pat Singhania University, Udaipur, Rajasthan

Mr. Muhammad Imran Khan, Universiti Teknologi PETRONAS, Malaysia

Dr. Natarajan Meghanathan, Jackson State University, Jackson, MS, USA

Mr. Mohd Zaki Bin Mas'ud, Universiti Teknikal Malaysia Melaka (UTeM), Malaysia

Prof. Dr. R. Geetharamani, Dept. of Computer Science and Eng., Rajalakshmi Engineering College, India

Dr. Smita Rajpal, Institute of Technology and Management, Gurgaon, India

Dr. S. Abdul Khader Jilani, University of Tabuk, Tabuk, Saudi Arabia

Mr. Syed Jamal Haider Zaidi, Bahria University, Pakistan

Dr. N. Devarajan, Government College of Technology,Coimbatore, Tamilnadu, INDIA

Mr. R. Jagadeesh Kannan, RMK Engineering College, India

Mr. Deo Prakash, Shri Mata Vaishno Devi University, India

Mr. Mohammad Abu Naser, Dept. of EEE, IUT, Gazipur, Bangladesh

Assist. Prof. Prasun Ghosal, Bengal Engineering and Science University, India

Mr. Md. Golam Kaosar, School of Engineering and Science, Victoria University, Melbourne City, Australia

Mr. R. Mahammad Shafi, Madanapalle Institute of Technology & Science, India

Dr. F.Sagayaraj Francis, Pondicherry Engineering College,India

Dr. Ajay Goel, HIET , Kaithal, India

Mr. Nayak Sunil Kashibarao, Bahirji Smarak Mahavidyalaya, India

Mr. Suhas J Manangi, Microsoft India

Dr. Kalyankar N. V., Yeshwant Mahavidyalaya, Nanded , India

Dr. K.D. Verma, S.V. College of Post graduate studies & Research, India

Dr. Amjad Rehman, University Technology Malaysia, Malaysia


Mr. Rachit Garg, L K College, Jalandhar, Punjab

Mr. J. William, M.A.M college of Engineering, Trichy, Tamilnadu,India

Prof. Jue-Sam Chou, Nanhua University, College of Science and Technology, Taiwan

Dr. Thorat S.B., Institute of Technology and Management, India

Mr. Ajay Prasad, Sir Padampat Singhania University, Udaipur, India

Dr. Kamaljit I. Lakhtaria, Atmiya Institute of Technology & Science, India

Mr. Syed Rafiul Hussain, Ahsanullah University of Science and Technology, Bangladesh

Mrs Fazeela Tunnisa, Najran University, Kingdom of Saudi Arabia

Mrs Kavita Taneja, Maharishi Markandeshwar University, Haryana, India

Mr. Maniyar Shiraz Ahmed, Najran University, Najran, KSA

Mr. Anand Kumar, AMC Engineering College, Bangalore

Dr. Rakesh Chandra Gangwar, Beant College of Engg. & Tech., Gurdaspur (Punjab) India

Dr. V V Rama Prasad, Sree Vidyanikethan Engineering College, India

Assist. Prof. Neetesh Kumar Gupta, Technocrats Institute of Technology, Bhopal (M.P.), India

Mr. Ashish Seth, Uttar Pradesh Technical University, Lucknow ,UP India

Dr. V V S S S Balaram, Sreenidhi Institute of Science and Technology, India

Mr Rahul Bhatia, Lingaya's Institute of Management and Technology, India

Prof. Niranjan Reddy. P, KITS , Warangal, India

Prof. Rakesh. Lingappa, Vijetha Institute of Technology, Bangalore, India

Dr. Mohammed Ali Hussain, Nimra College of Engineering & Technology, Vijayawada, A.P., India

Dr. A.Srinivasan, MNM Jain Engineering College, Rajiv Gandhi Salai, Thorapakkam, Chennai

Mr. Rakesh Kumar, M.M. University, Mullana, Ambala, India

Dr. Lena Khaled, Zarqa Private University, Aman, Jordon

Ms. Supriya Kapoor, Patni/Lingaya's Institute of Management and Tech., India

Dr. Tossapon Boongoen , Aberystwyth University, UK

Dr . Bilal Alatas, Firat University, Turkey

Assist. Prof. Jyoti Praaksh Singh , Academy of Technology, India

Dr. Ritu Soni, GNG College, India

Dr . Mahendra Kumar , Sagar Institute of Research & Technology, Bhopal, India.

Dr. Binod Kumar, Lakshmi Narayan College of Tech.(LNCT)Bhopal India

Dr. Muzhir Shaban Al-Ani, Amman Arab University Amman – Jordan

Dr. T.C. Manjunath , ATRIA Institute of Tech, India

Mr. Muhammad Zakarya, COMSATS Institute of Information Technology (CIIT), Pakistan

Assist. Prof. Harmunish Taneja, M. M. University, India

Dr. Chitra Dhawale , SICSR, Model Colony, Pune, India

Mrs Sankari Muthukaruppan, Nehru Institute of Engineering and Technology, Anna University, India

Mr. Aaqif Afzaal Abbasi, National University Of Sciences And Technology, Islamabad

Prof. Ashutosh Kumar Dubey, Trinity Institute of Technology and Research Bhopal, India

Mr. G. Appasami, Dr. Pauls Engineering College, India

Mr. M Yasin, National University of Science and Tech, karachi (NUST), Pakistan

Mr. Yaser Miaji, University Utara Malaysia, Malaysia

Mr. Shah Ahsanul Haque, International Islamic University Chittagong (IIUC), Bangladesh


Prof. (Dr) Syed Abdul Sattar, Royal Institute of Technology & Science, India

Dr. S. Sasikumar, Roever Engineering College

Assist. Prof. Monit Kapoor, Maharishi Markandeshwar University, India

Mr. Nwaocha Vivian O, National Open University of Nigeria

Dr. M. S. Vijaya, GR Govindarajulu School of Applied Computer Technology, India

Assist. Prof. Chakresh Kumar, Manav Rachna International University, India

Mr. Kunal Chadha , R&D Software Engineer, Gemalto, Singapore

Mr. Mueen Uddin, Universiti Teknologi Malaysia, UTM , Malaysia

Dr. Dhuha Basheer abdullah, Mosul university, Iraq

Mr. S. Audithan, Annamalai University, India

Prof. Vijay K Chaudhari, Technocrats Institute of Technology , India

Associate Prof. Mohd Ilyas Khan, Technocrats Institute of Technology , India

Dr. Vu Thanh Nguyen, University of Information Technology, HoChiMinh City, VietNam

Assist. Prof. Anand Sharma, MITS, Lakshmangarh, Sikar, Rajasthan, India

Prof. T V Narayana Rao, HITAM Engineering college, Hyderabad

Mr. Deepak Gour, Sir Padampat Singhania University, India

Assist. Prof. Amutharaj Joyson, Kalasalingam University, India

Mr. Ali Balador, Islamic Azad University, Iran

Mr. Mohit Jain, Maharaja Surajmal Institute of Technology, India

Mr. Dilip Kumar Sharma, GLA Institute of Technology & Management, India

Dr. Debojyoti Mitra, Sir padampat Singhania University, India

Dr. Ali Dehghantanha, Asia-Pacific University College of Technology and Innovation, Malaysia

Mr. Zhao Zhang, City University of Hong Kong, China

Prof. S.P. Setty, A.U. College of Engineering, India

Prof. Patel Rakeshkumar Kantilal, Sankalchand Patel College of Engineering, India

Mr. Biswajit Bhowmik, Bengal College of Engineering & Technology, India

Mr. Manoj Gupta, Apex Institute of Engineering & Technology, India

Assist. Prof. Ajay Sharma, Raj Kumar Goel Institute Of Technology, India

Assist. Prof. Ramveer Singh, Raj Kumar Goel Institute of Technology, India

Dr. Hanan Elazhary, Electronics Research Institute, Egypt

Dr. Hosam I. Faiq, USM, Malaysia

Prof. Dipti D. Patil, MAEER’s MIT College of Engg. & Tech, Pune, India

Assist. Prof. Devendra Chack, BCT Kumaon engineering College Dwarahat Almora, India

Prof. Manpreet Singh, M. M. Engg. College, M. M. University, India

Assist. Prof. M. Sadiq ali Khan, University of Karachi, Pakistan

Mr. Prasad S. Halgaonkar, MIT - College of Engineering, Pune, India

Dr. Imran Ghani, Universiti Teknologi Malaysia, Malaysia

Prof. Varun Kumar Kakar, Kumaon Engineering College, Dwarahat, India

Assist. Prof. Nisheeth Joshi, Apaji Institute, Banasthali University, Rajasthan, India

Associate Prof. Kunwar S. Vaisla, VCT Kumaon Engineering College, India

Prof Anupam Choudhary, Bhilai School Of Engg.,Bhilai (C.G.),India

Mr. Divya Prakash Shrivastava, Al Jabal Al garbi University, Zawya, Libya


Associate Prof. Dr. V. Radha, Avinashilingam Deemed university for women, Coimbatore.

Dr. Kasarapu Ramani, JNT University, Anantapur, India

Dr. Anuraag Awasthi, Jayoti Vidyapeeth Womens University, India

Dr. C G Ravichandran, R V S College of Engineering and Technology, India

Dr. Mohamed A. Deriche, King Fahd University of Petroleum and Minerals, Saudi Arabia

Mr. Abbas Karimi, Universiti Putra Malaysia, Malaysia

Mr. Amit Kumar, Jaypee University of Engg. and Tech., India

Dr. Nikolai Stoianov, Defense Institute, Bulgaria

Assist. Prof. S. Ranichandra, KSR College of Arts and Science, Tiruchencode

Mr. T.K.P. Rajagopal, Diamond Horse International Pvt Ltd, India

Dr. Md. Ekramul Hamid, Rajshahi University, Bangladesh

Mr. Hemanta Kumar Kalita , TATA Consultancy Services (TCS), India

Dr. Messaouda Azzouzi, Ziane Achour University of Djelfa, Algeria

Prof. (Dr.) Juan Jose Martinez Castillo, "Gran Mariscal de Ayacucho" University and Acantelys research

Group, Venezuela

Dr. Jatinderkumar R. Saini, Narmada College of Computer Application, India

Dr. Babak Bashari Rad, University Technology of Malaysia, Malaysia

Dr. Nighat Mir, Effat University, Saudi Arabia

Prof. (Dr.) G.M.Nasira, Sasurie College of Engineering, India

Mr. Varun Mittal, Gemalto Pte Ltd, Singapore

Assist. Prof. Mrs P. Banumathi, Kathir College Of Engineering, Coimbatore

Assist. Prof. Quan Yuan, University of Wisconsin-Stevens Point, US

Dr. Pranam Paul, Narula Institute of Technology, Agarpara, West Bengal, India

Assist. Prof. J. Ramkumar, V.L.B Janakiammal college of Arts & Science, India

Mr. P. Sivakumar, Anna university, Chennai, India

Mr. Md. Humayun Kabir Biswas, King Khalid University, Kingdom of Saudi Arabia

Mr. Mayank Singh, J.P. Institute of Engg & Technology, Meerut, India

HJ. Kamaruzaman Jusoff, Universiti Putra Malaysia

Mr. Nikhil Patrick Lobo, CADES, India

Dr. Amit Wason, Rayat-Bahra Institute of Engineering & Boi-Technology, India

Dr. Rajesh Shrivastava, Govt. Benazir Science & Commerce College, Bhopal, India

Assist. Prof. Vishal Bharti, DCE, Gurgaon

Mrs. Sunita Bansal, Birla Institute of Technology & Science, India

Dr. R. Sudhakar, Dr.Mahalingam college of Engineering and Technology, India

Dr. Amit Kumar Garg, Shri Mata Vaishno Devi University, Katra(J&K), India

Assist. Prof. Raj Gaurang Tiwari, AZAD Institute of Engineering and Technology, India

Mr. Hamed Taherdoost, Tehran, Iran

Mr. Amin Daneshmand Malayeri, YRC, IAU, Malayer Branch, Iran

Mr. Shantanu Pal, University of Calcutta, India

Dr. Terry H. Walcott, E-Promag Consultancy Group, United Kingdom

Dr. Ezekiel U OKIKE, University of Ibadan, Nigeria

Mr. P. Mahalingam, Caledonian College of Engineering, Oman


Dr. Mahmoud M. A. Abd Ellatif, Mansoura University, Egypt

Prof. Kunwar S. Vaisla, BCT Kumaon Engineering College, India

Prof. Mahesh H. Panchal, Kalol Institute of Technology & Research Centre, India

Mr. Muhammad Asad, Technical University of Munich, Germany

Mr. AliReza Shams Shafigh, Azad Islamic university, Iran

Prof. S. V. Nagaraj, RMK Engineering College, India

Mr. Ashikali M Hasan, Senior Researcher, CelNet security, India

Dr. Adnan Shahid Khan, University Technology Malaysia, Malaysia

Mr. Prakash Gajanan Burade, Nagpur University/ITM college of engg, Nagpur, India

Dr. Jagdish B.Helonde, Nagpur University/ITM college of engg, Nagpur, India

Professor, Doctor BOUHORMA Mohammed, Univertsity Abdelmalek Essaadi, Morocco

Mr. K. Thirumalaivasan, Pondicherry Engg. College, India

Mr. Umbarkar Anantkumar Janardan, Walchand College of Engineering, India

Mr. Ashish Chaurasia, Gyan Ganga Institute of Technology & Sciences, India

Mr. Sunil Taneja, Kurukshetra University, India

Mr. Fauzi Adi Rafrastara, Dian Nuswantoro University, Indonesia

Dr. Yaduvir Singh, Thapar University, India

Dr. Ioannis V. Koskosas, University of Western Macedonia, Greece

Dr. Vasantha Kalyani David, Avinashilingam University for women, Coimbatore

Dr. Ahmed Mansour Manasrah, Universiti Sains Malaysia, Malaysia

Miss. Nazanin Sadat Kazazi, University Technology Malaysia, Malaysia

Mr. Saeed Rasouli Heikalabad, Islamic Azad University - Tabriz Branch, Iran

Assoc. Prof. Dhirendra Mishra, SVKM's NMIMS University, India

Prof. Shapoor Zarei, UAE Inventors Association, UAE

Prof. B.Raja Sarath Kumar, Lenora College of Engineering, India

Dr. Bashir Alam, Jamia millia Islamia, Delhi, India

Prof. Anant J Umbarkar, Walchand College of Engg., India

Assist. Prof. B. Bharathi, Sathyabama University, India

Dr. Fokrul Alom Mazarbhuiya, King Khalid University, Saudi Arabia

Prof. T.S.Jeyali Laseeth, Anna University of Technology, Tirunelveli, India

Dr. M. Balraju, Jawahar Lal Nehru Technological University Hyderabad, India

Dr. Vijayalakshmi M. N., R.V.College of Engineering, Bangalore

Prof. Walid Moudani, Lebanese University, Lebanon

Dr. Saurabh Pal, VBS Purvanchal University, Jaunpur, India

Associate Prof. Suneet Chaudhary, Dehradun Institute of Technology, India

Associate Prof. Dr. Manuj Darbari, BBD University, India

Ms. Prema Selvaraj, K.S.R College of Arts and Science, India

Assist. Prof. Ms.S.Sasikala, KSR College of Arts & Science, India

Mr. Sukhvinder Singh Deora, NC Institute of Computer Sciences, India

Dr. Abhay Bansal, Amity School of Engineering & Technology, India

Ms. Sumita Mishra, Amity School of Engineering and Technology, India

Professor S. Viswanadha Raju, JNT University Hyderabad, India


Mr. Asghar Shahrzad Khashandarag, Islamic Azad University Tabriz Branch, India

Mr. Manoj Sharma, Panipat Institute of Engg. & Technology, India

Mr. Shakeel Ahmed, King Faisal University, Saudi Arabia

Dr. Mohamed Ali Mahjoub, Institute of Engineer of Monastir, Tunisia

Mr. Adri Jovin J.J., SriGuru Institute of Technology, India

Dr. Sukumar Senthilkumar, Universiti Sains Malaysia, Malaysia

Mr. Rakesh Bharati, Dehradun Institute of Technology Dehradun, India

Mr. Shervan Fekri Ershad, Shiraz International University, Iran

Mr. Md. Safiqul Islam, Daffodil International University, Bangladesh

Mr. Mahmudul Hasan, Daffodil International University, Bangladesh

Prof. Mandakini Tayade, UIT, RGTU, Bhopal, India

Ms. Sarla More, UIT, RGTU, Bhopal, India

Mr. Tushar Hrishikesh Jaware, R.C. Patel Institute of Technology, Shirpur, India

Ms. C. Divya, Dr G R Damodaran College of Science, Coimbatore, India

Mr. Fahimuddin Shaik, Annamacharya Institute of Technology & Sciences, India

Dr. M. N. Giri Prasad, JNTUCE,Pulivendula, A.P., India

Assist. Prof. Chintan M Bhatt, Charotar University of Science And Technology, India

Prof. Sahista Machchhar, Marwadi Education Foundation's Group of institutions, India

Assist. Prof. Navnish Goel, S. D. College Of Enginnering & Technology, India

Mr. Khaja Kamaluddin, Sirt University, Sirt, Libya

Mr. Mohammad Zaidul Karim, Daffodil International, Bangladesh

Mr. M. Vijayakumar, KSR College of Engineering, Tiruchengode, India

Mr. S. A. Ahsan Rajon, Khulna University, Bangladesh

Dr. Muhammad Mohsin Nazir, LCW University Lahore, Pakistan

Mr. Mohammad Asadul Hoque, University of Alabama, USA

Mr. P.V.Sarathchand, Indur Institute of Engineering and Technology, India

Mr. Durgesh Samadhiya, Chung Hua University, Taiwan

Dr Venu Kuthadi, University of Johannesburg, Johannesburg, RSA

Dr. (Er) Jasvir Singh, Guru Nanak Dev University, Amritsar, Punjab, India

Mr. Jasmin Cosic, Min. of the Interior of Una-sana canton, B&H, Bosnia and Herzegovina

Dr S. Rajalakshmi, Botho College, South Africa

Dr. Mohamed Sarrab, De Montfort University, UK

Mr. Basappa B. Kodada, Canara Engineering College, India

Assist. Prof. K. Ramana, Annamacharya Institute of Technology and Sciences, India

Dr. Ashu Gupta, Apeejay Institute of Management, Jalandhar, India

Assist. Prof. Shaik Rasool, Shadan College of Engineering & Technology, India

Assist. Prof. K. Suresh, Annamacharya Institute of Tech & Sci. Rajampet, AP, India

Dr . G. Singaravel, K.S.R. College of Engineering, India

Dr B. G. Geetha, K.S.R. College of Engineering, India

Assist. Prof. Kavita Choudhary, ITM University, Gurgaon

Dr. Mehrdad Jalali, Azad University, Mashhad, Iran

Megha Goel, Shamli Institute of Engineering and Technology, Shamli, India


Mr. Chi-Hua Chen, Institute of Information Management, National Chiao-Tung University, Taiwan (R.O.C.)

Assoc. Prof. A. Rajendran, RVS College of Engineering and Technology, India

Assist. Prof. S. Jaganathan, RVS College of Engineering and Technology, India

Assoc. Prof. (Dr.) A S N Chakravarthy, JNTUK University College of Engineering Vizianagaram (State

University)

Assist. Prof. Deepshikha Patel, Technocrat Institute of Technology, India

Assist. Prof. Maram Balajee, GMRIT, India

Assist. Prof. Monika Bhatnagar, TIT, India

Prof. Gaurang Panchal, Charotar University of Science & Technology, India

Prof. Anand K. Tripathi, Computer Society of India

Prof. Jyoti Chaudhary, High Performance Computing Research Lab, India

Assist. Prof. Supriya Raheja, ITM University, India

Dr. Pankaj Gupta, Microsoft Corporation, U.S.A.

Assist. Prof. Panchamukesh Chandaka, Hyderabad Institute of Tech. & Management, India

Prof. Mohan H.S, SJB Institute Of Technology, India

Mr. Hossein Malekinezhad, Islamic Azad University, Iran

Mr. Zatin Gupta, Universti Malaysia, Malaysia

Assist. Prof. Amit Chauhan, Phonics Group of Institutions, India

Assist. Prof. Ajal A. J., METS School Of Engineering, India

Mrs. Omowunmi Omobola Adeyemo, University of Ibadan, Nigeria

Dr. Bharat Bhushan Agarwal, I.F.T.M. University, India

Md. Nazrul Islam, University of Western Ontario, Canada

Tushar Kanti, L.N.C.T, Bhopal, India

Er. Aumreesh Kumar Saxena, SIRTs College Bhopal, India

Mr. Mohammad Monirul Islam, Daffodil International University, Bangladesh

Dr. Kashif Nisar, University Utara Malaysia, Malaysia

Dr. Wei Zheng, Rutgers Univ/ A10 Networks, USA

Associate Prof. Rituraj Jain, Vyas Institute of Engg & Tech, Jodhpur – Rajasthan

Assist. Prof. Apoorvi Sood, I.T.M. University, India

Dr. Kayhan Zrar Ghafoor, University Technology Malaysia, Malaysia

Mr. Swapnil Soner, Truba Institute College of Engineering & Technology, Indore, India

Ms. Yogita Gigras, I.T.M. University, India

Associate Prof. Neelima Sadineni, Pydha Engineering College, India Pydha Engineering College

Assist. Prof. K. Deepika Rani, HITAM, Hyderabad

Ms. Shikha Maheshwari, Jaipur Engineering College & Research Centre, India

Prof. Dr V S Giridhar Akula, Avanthi's Scientific Tech. & Research Academy, Hyderabad

Prof. Dr.S.Saravanan, Muthayammal Engineering College, India

Mr. Mehdi Golsorkhatabar Amiri, Islamic Azad University, Iran

Prof. Amit Sadanand Savyanavar, MITCOE, Pune, India

Assist. Prof. P.Oliver Jayaprakash, Anna University,Chennai

Assist. Prof. Ms. Sujata, ITM University, Gurgaon, India

Dr. Asoke Nath, St. Xavier's College, India


Mr. Masoud Rafighi, Islamic Azad University, Iran

Assist. Prof. RamBabu Pemula, NIMRA College of Engineering & Technology, India

Assist. Prof. Ms Rita Chhikara, ITM University, Gurgaon, India

Mr. Sandeep Maan, Government Post Graduate College, India

Prof. Dr. S. Muralidharan, Mepco Schlenk Engineering College, India

Associate Prof. T.V.Sai Krishna, QIS College of Engineering and Technology, India

Mr. R. Balu, Bharathiar University, Coimbatore, India

Assist. Prof. Shekhar. R, Dr.SM College of Engineering, India

Prof. P. Senthilkumar, Vivekanandha Institue of Engineering and Techology for Woman, India

Mr. M. Kamarajan, PSNA College of Engineering & Technology, India

Dr. Angajala Srinivasa Rao, Jawaharlal Nehru Technical University, India

Assist. Prof. C. Venkatesh, A.I.T.S, Rajampet, India

Mr. Afshin Rezakhani Roozbahani, Ayatollah Boroujerdi University, Iran

Mr. Laxmi chand, SCTL, Noida, India

Dr. Dr. Abdul Hannan, Vivekanand College, Aurangabad

Prof. Mahesh Panchal, KITRC, Gujarat

Dr. A. Subramani, K.S.R. College of Engineering, Tiruchengode

Assist. Prof. Prakash M, Rajalakshmi Engineering College, Chennai, India

Assist. Prof. Akhilesh K Sharma, Sir Padampat Singhania University, India

Ms. Varsha Sahni, Guru Nanak Dev Engineering College, Ludhiana, India

Associate Prof. Trilochan Rout, NM Institute of Engineering and Technlogy, India

Mr. Srikanta Kumar Mohapatra, NMIET, Orissa, India

Mr. Waqas Haider Bangyal, Iqra University Islamabad, Pakistan

Dr. S. Vijayaragavan, Christ College of Engineering and Technology, Pondicherry, India

Prof. Elboukhari Mohamed, University Mohammed First, Oujda, Morocco

Dr. Muhammad Asif Khan, King Faisal University, Saudi Arabia

Dr. Nagy Ramadan Darwish Omran, Cairo University, Egypt.

Assistant Prof. Anand Nayyar, KCL Institute of Management and Technology, India

Mr. G. Premsankar, Ericcson, India

Assist. Prof. T. Hemalatha, VELS University, India

Prof. Tejaswini Apte, University of Pune, India

Dr. Edmund Ng Giap Weng, Universiti Malaysia Sarawak, Malaysia

Mr. Mahdi Nouri, Iran University of Science and Technology, Iran

Associate Prof. S. Asif Hussain, Annamacharya Institute of technology & Sciences, India

Mrs. Kavita Pabreja, Maharaja Surajmal Institute (an affiliate of GGSIP University), India

Mr. Vorugunti Chandra Sekhar, DA-IICT, India

Mr. Muhammad Najmi Ahmad Zabidi, Universiti Teknologi Malaysia, Malaysia

Dr. Aderemi A. Atayero, Covenant University, Nigeria

Assist. Prof. Osama Sohaib, Balochistan University of Information Technology, Pakistan

Assist. Prof. K. Suresh, Annamacharya Institute of Technology and Sciences, India

Mr. Hassen Mohammed Abduallah Alsafi, International Islamic University Malaysia (IIUM) Malaysia

Mr. Robail Yasrab, Virtual University of Pakistan, Pakistan


Mr. R. Balu, Bharathiar University, Coimbatore, India

Prof. Anand Nayyar, KCL Institute of Management and Technology, Jalandhar

Assoc. Prof. Vivek S Deshpande, MIT College of Engineering, India

Prof. K. Saravanan, Anna university Coimbatore, India

Dr. Ravendra Singh, MJP Rohilkhand University, Bareilly, India

Mr. V. Mathivanan, IBRA College of Technology, Sultanate of OMAN

Assoc. Prof. S. Asif Hussain, AITS, India

Assist. Prof. C. Venkatesh, AITS, India

Mr. Sami Ulhaq, SZABIST Islamabad, Pakistan

Dr. B. Justus Rabi, Institute of Science & Technology, India

Mr. Anuj Kumar Yadav, Dehradun Institute of technology, India

Mr. Alejandro Mosquera, University of Alicante, Spain

Assist. Prof. Arjun Singh, Sir Padampat Singhania University (SPSU), Udaipur, India

Dr. Smriti Agrawal, JB Institute of Engineering and Technology, Hyderabad

Assist. Prof. Swathi Sambangi, Visakha Institute of Engineering and Technology, India

Ms. Prabhjot Kaur, Guru Gobind Singh Indraprastha University, India

Mrs. Samaher AL-Hothali, Yanbu University College, Saudi Arabia

Prof. Rajneeshkaur Bedi, MIT College of Engineering, Pune, India

Mr. Hassen Mohammed Abduallah Alsafi, International Islamic University Malaysia (IIUM)

Dr. Wei Zhang, Amazon.com, Seattle, WA, USA

Mr. B. Santhosh Kumar, C S I College of Engineering, Tamil Nadu

Dr. K. Reji Kumar, , N S S College, Pandalam, India

Assoc. Prof. K. Seshadri Sastry, EIILM University, India

Mr. Kai Pan, UNC Charlotte, USA

Mr. Ruikar Sachin, SGGSIET, India

Prof. (Dr.) Vinodani Katiyar, Sri Ramswaroop Memorial University, India

Assoc. Prof., M. Giri, Sreenivasa Institute of Technology and Management Studies, India

Assoc. Prof. Labib Francis Gergis, Misr Academy for Engineering and Technology (MET), Egypt

Assist. Prof. Amanpreet Kaur, ITM University, India

Assist. Prof. Anand Singh Rajawat, Shri Vaishnav Institute of Technology & Science, Indore

Mrs. Hadeel Saleh Haj Aliwi, Universiti Sains Malaysia (USM), Malaysia

Dr. Abhay Bansal, Amity University, India

Dr. Mohammad A. Mezher, Fahad Bin Sultan University, KSA

Assist. Prof. Nidhi Arora, M.C.A. Institute, India

Prof. Dr. P. Suresh, Karpagam College of Engineering, Coimbatore, India

Dr. Kannan Balasubramanian, Mepco Schlenk Engineering College, India

Dr. S. Sankara Gomathi, Panimalar Engineering college, India

Prof. Anil kumar Suthar, Gujarat Technological University, L.C. Institute of Technology, India

Assist. Prof. R. Hubert Rajan, NOORUL ISLAM UNIVERSITY, India

Assist. Prof. Dr. Jyoti Mahajan, College of Engineering & Technology

Assist. Prof. Homam Reda El-Taj, College of Network Engineering, Saudi Arabia & Malaysia

Mr. Bijan Paul, Shahjalal University of Science & Technology, Bangladesh


Assoc. Prof. Dr. Ch V Phani Krishna, KL University, India

Dr. Vishal Bhatnagar, Ambedkar Institute of Advanced Communication Technologies & Research, India

Dr. Lamri LAOUAMER, Al Qassim University, Dept. Info. Systems & European University of Brittany, Dept.

Computer Science, UBO, Brest, France

Prof. Ashish Babanrao Sasankar, G.H.Raisoni Institute Of Information Technology, India

Prof. Pawan Kumar Goel, Shamli Institute of Engineering and Technology, India

Mr. Ram Kumar Singh, S.V Subharti University, India

Assistant Prof. Sunish Kumar O S, Amaljyothi College of Engineering, India

Dr Sanjay Bhargava, Banasthali University, India

Mr. Pankaj S. Kulkarni, AVEW's Shatabdi Institute of Technology, India

Mr. Roohollah Etemadi, Islamic Azad University, Iran

Mr. Oloruntoyin Sefiu Taiwo, Emmanuel Alayande College Of Education, Nigeria

Mr. Sumit Goyal, National Dairy Research Institute, India

Mr Jaswinder Singh Dilawari, Geeta Engineering College, India

Prof. Raghuraj Singh, Harcourt Butler Technological Institute, Kanpur

Dr. S.K. Mahendran, Anna University, Chennai, India

Dr. Amit Wason, Hindustan Institute of Technology & Management, Punjab

Dr. Ashu Gupta, Apeejay Institute of Management, India

Assist. Prof. D. Asir Antony Gnana Singh, M.I.E.T Engineering College, India

Mrs Mina Farmanbar, Eastern Mediterranean University, Famagusta, North Cyprus

Mr. Maram Balajee, GMR Institute of Technology, India

Mr. Moiz S. Ansari, Isra University, Hyderabad, Pakistan

Mr. Adebayo, Olawale Surajudeen, Federal University of Technology Minna, Nigeria

Mr. Jasvir Singh, University College Of Engg., India

Mr. Vivek Tiwari, MANIT, Bhopal, India

Assoc. Prof. R. Navaneethakrishnan, Bharathiyar College of Engineering and Technology, India

Mr. Somdip Dey, St. Xavier's College, Kolkata, India

Mr. Souleymane Balla-Arabé, Xi’an University of Electronic Science and Technology, China

Mr. Mahabub Alam, Rajshahi University of Engineering and Technology, Bangladesh

Mr. Sathyapraksh P., S.K.P Engineering College, India

Dr. N. Karthikeyan, SNS College of Engineering, Anna University, India

Dr. Binod Kumar, JSPM's, Jayawant Technical Campus, Pune, India

Assoc. Prof. Dinesh Goyal, Suresh Gyan Vihar University, India

Mr. Md. Abdul Ahad, K L University, India

Mr. Vikas Bajpai, The LNM IIT, India

Dr. Manish Kumar Anand, Salesforce (R & D Analytics), San Francisco, USA

Assist. Prof. Dheeraj Murari, Kumaon Engineering College, India

Assoc. Prof. Dr. A. Muthukumaravel, VELS University, Chennai

Mr. A. Siles Balasingh, St.Joseph University in Tanzania, Tanzania

Mr. Ravindra Daga Badgujar, R C Patel Institute of Technology, India

Dr. Preeti Khanna, SVKM’s NMIMS, School of Business Management, India

Mr. Kumar Dayanand, Cambridge Institute of Technology, India


Dr. Syed Asif Ali, SMI University Karachi, Pakistan

Prof. Pallvi Pandit, Himachal Pradeh University, India

Mr. Ricardo Verschueren, University of Gloucestershire, UK

Assist. Prof. Mamta Juneja, University Institute of Engineering and Technology, Panjab University, India

Assoc. Prof. P. Surendra Varma, NRI Institute of Technology, JNTU Kakinada, India

Assist. Prof. Gaurav Shrivastava, RGPV / SVITS Indore, India

Dr. S. Sumathi, Anna University, India

Assist. Prof. Ankita M. Kapadia, Charotar University of Science and Technology, India

Mr. Deepak Kumar, Indian Institute of Technology (BHU), India

Dr. Dr. Rajan Gupta, GGSIP University, New Delhi, India

Assist. Prof M. Anand Kumar, Karpagam University, Coimbatore, India

Mr. Mr Arshad Mansoor, Pakistan Aeronautical Complex

Mr. Kapil Kumar Gupta, Ansal Institute of Technology and Management, India

Dr. Neeraj Tomer, SINE International Institute of Technology, Jaipur, India

Assist. Prof. Trunal J. Patel, C.G.Patel Institute of Technology, Uka Tarsadia University, Bardoli, Surat

Mr. Sivakumar, Codework solutions, India

Mr. Mohammad Sadegh Mirzaei, PGNR Company, Iran

Dr. Gerard G. Dumancas, Oklahoma Medical Research Foundation, USA

Mr. Varadala Sridhar, Varadhaman College Engineering College, Affiliated To JNTU, Hyderabad

Assist. Prof. Manoj Dhawan, SVITS, Indore

Assoc. Prof. Chitreshh Banerjee, Suresh Gyan Vihar University, Jaipur, India

Dr. S. Santhi, SCSVMV University, India

Mr. Davood Mohammadi Souran, Ministry of Energy of Iran, Iran

Mr. Shamim Ahmed, Bangladesh University of Business and Technology, Bangladesh

Mr. Sandeep Reddivari, Mississippi State University, USA

Assoc. Prof. Ousmane Thiare, Gaston Berger University, Senegal

Dr. Hazra Imran, Athabasca University, Canada

Dr. Setu Kumar Chaturvedi, Technocrats Institute of Technology, Bhopal, India

Mr. Mohd Dilshad Ansari, Jaypee University of Information Technology, India

Ms. Jaspreet Kaur, Distance Education LPU, India

Dr. D. Nagarajan, Salalah College of Technology, Sultanate of Oman

Dr. K.V.N.R.Sai Krishna, S.V.R.M. College, India

Mr. Himanshu Pareek, Center for Development of Advanced Computing (CDAC), India

Mr. Khaldi Amine, Badji Mokhtar University, Algeria

Mr. Mohammad Sadegh Mirzaei, Scientific Applied University, Iran

Assist. Prof. Khyati Chaudhary, Ram-eesh Institute of Engg. & Technology, India

Mr. Sanjay Agal, Pacific College of Engineering Udaipur, India

Mr. Abdul Mateen Ansari, King Khalid University, Saudi Arabia

Dr. H.S. Behera, Veer Surendra Sai University of Technology (VSSUT), India

Dr. Shrikant Tiwari, Shri Shankaracharya Group of Institutions (SSGI), India

Prof. Ganesh B. Regulwar, Shri Shankarprasad Agnihotri College of Engg, India

Prof. Pinnamaneni Bhanu Prasad, Matrix vision GmbH, Germany


Dr. Shrikant Tiwari, Shri Shankaracharya Technical Campus (SSTC), India

Dr. Siddesh G.K., : Dayananada Sagar College of Engineering, Bangalore, India

Mr. Nadir Bouchama, CERIST Research Center, Algeria

Dr. R. Sathishkumar, Sri Venkateswara College of Engineering, India

Assistant Prof (Dr.) Mohamed Moussaoui, Abdelmalek Essaadi University, Morocco

Dr. S. Malathi, Panimalar Engineering College, Chennai, India

Dr. V. Subedha, Panimalar Institute of Technology, Chennai, India

Dr. Prashant Panse, Swami Vivekanand College of Engineering, Indore, India

Dr. Hamza Aldabbas, Al-Balqa’a Applied University, Jordan

Dr. G. Rasitha Banu, Vel's University, Chennai

Dr. V. D. Ambeth Kumar, Panimalar Engineering College, Chennai

Prof. Anuranjan Misra, Bhagwant Institute of Technology, Ghaziabad, India

Ms. U. Sinthuja, PSG college of arts &science, India

Mr. Ehsan Saradar Torshizi, Urmia University, Iran

Mr. Shamneesh Sharma, APG Shimla University, Shimla (H.P.), India

CALL FOR PAPERS International Journal of Computer Science and Information Security

IJCSIS 2014 ISSN: 1947-5500

http://sites.google.com/site/ijcsis/ International Journal Computer Science and Information Security, IJCSIS, is the premier scholarly venue in the areas of computer science and security issues. IJCSIS 2011 will provide a high profile, leading edge platform for researchers and engineers alike to publish state-of-the-art research in the respective fields of information technology and communication security. The journal will feature a diverse mixture of publication articles including core and applied computer science related topics. Authors are solicited to contribute to the special issue by submitting articles that illustrate research results, projects, surveying works and industrial experiences that describe significant advances in the following areas, but are not limited to. Submissions may span a broad range of topics, e.g.: Track A: Security Access control, Anonymity, Audit and audit reduction & Authentication and authorization, Applied cryptography, Cryptanalysis, Digital Signatures, Biometric security, Boundary control devices, Certification and accreditation, Cross-layer design for security, Security & Network Management, Data and system integrity, Database security, Defensive information warfare, Denial of service protection, Intrusion Detection, Anti-malware, Distributed systems security, Electronic commerce, E-mail security, Spam, Phishing, E-mail fraud, Virus, worms, Trojan Protection, Grid security, Information hiding and watermarking & Information survivability, Insider threat protection, Integrity Intellectual property protection, Internet/Intranet Security, Key management and key recovery, Language-based security, Mobile and wireless security, Mobile, Ad Hoc and Sensor Network Security, Monitoring and surveillance, Multimedia security ,Operating system security, Peer-to-peer security, Performance Evaluations of Protocols & Security Application, Privacy and data protection, Product evaluation criteria and compliance, Risk evaluation and security certification, Risk/vulnerability assessment, Security & Network Management, Security Models & protocols, Security threats & countermeasures (DDoS, MiM, Session Hijacking, Replay attack etc,), Trusted computing, Ubiquitous Computing Security, Virtualization security, VoIP security, Web 2.0 security, Submission Procedures, Active Defense Systems, Adaptive Defense Systems, Benchmark, Analysis and Evaluation of Security Systems, Distributed Access Control and Trust Management, Distributed Attack Systems and Mechanisms, Distributed Intrusion Detection/Prevention Systems, Denial-of-Service Attacks and Countermeasures, High Performance Security Systems, Identity Management and Authentication, Implementation, Deployment and Management of Security Systems, Intelligent Defense Systems, Internet and Network Forensics, Large-scale Attacks and Defense, RFID Security and Privacy, Security Architectures in Distributed Network Systems, Security for Critical Infrastructures, Security for P2P systems and Grid Systems, Security in E-Commerce, Security and Privacy in Wireless Networks, Secure Mobile Agents and Mobile Code, Security Protocols, Security Simulation and Tools, Security Theory and Tools, Standards and Assurance Methods, Trusted Computing, Viruses, Worms, and Other Malicious Code, World Wide Web Security, Novel and emerging secure architecture, Study of attack strategies, attack modeling, Case studies and analysis of actual attacks, Continuity of Operations during an attack, Key management, Trust management, Intrusion detection techniques, Intrusion response, alarm management, and correlation analysis, Study of tradeoffs between security and system performance, Intrusion tolerance systems, Secure protocols, Security in wireless networks (e.g. mesh networks, sensor networks, etc.), Cryptography and Secure Communications, Computer Forensics, Recovery and Healing, Security Visualization, Formal Methods in Security, Principles for Designing a Secure Computing System, Autonomic Security, Internet Security, Security in Health Care Systems, Security Solutions Using Reconfigurable Computing, Adaptive and Intelligent Defense Systems, Authentication and Access control, Denial of service attacks and countermeasures, Identity, Route and

Location Anonymity schemes, Intrusion detection and prevention techniques, Cryptography, encryption algorithms and Key management schemes, Secure routing schemes, Secure neighbor discovery and localization, Trust establishment and maintenance, Confidentiality and data integrity, Security architectures, deployments and solutions, Emerging threats to cloud-based services, Security model for new services, Cloud-aware web service security, Information hiding in Cloud Computing, Securing distributed data storage in cloud, Security, privacy and trust in mobile computing systems and applications, Middleware security & Security features: middleware software is an asset on its own and has to be protected, interaction between security-specific and other middleware features, e.g., context-awareness, Middleware-level security monitoring and measurement: metrics and mechanisms for quantification and evaluation of security enforced by the middleware, Security co-design: trade-off and co-design between application-based and middleware-based security, Policy-based management: innovative support for policy-based definition and enforcement of security concerns, Identification and authentication mechanisms: Means to capture application specific constraints in defining and enforcing access control rules, Middleware-oriented security patterns: identification of patterns for sound, reusable security, Security in aspect-based middleware: mechanisms for isolating and enforcing security aspects, Security in agent-based platforms: protection for mobile code and platforms, Smart Devices: Biometrics, National ID cards, Embedded Systems Security and TPMs, RFID Systems Security, Smart Card Security, Pervasive Systems: Digital Rights Management (DRM) in pervasive environments, Intrusion Detection and Information Filtering, Localization Systems Security (Tracking of People and Goods), Mobile Commerce Security, Privacy Enhancing Technologies, Security Protocols (for Identification and Authentication, Confidentiality and Privacy, and Integrity), Ubiquitous Networks: Ad Hoc Networks Security, Delay-Tolerant Network Security, Domestic Network Security, Peer-to-Peer Networks Security, Security Issues in Mobile and Ubiquitous Networks, Security of GSM/GPRS/UMTS Systems, Sensor Networks Security, Vehicular Network Security, Wireless Communication Security: Bluetooth, NFC, WiFi, WiMAX, WiMedia, others This Track will emphasize the design, implementation, management and applications of computer communications, networks and services. Topics of mostly theoretical nature are also welcome, provided there is clear practical potential in applying the results of such work. Track B: Computer Science Broadband wireless technologies: LTE, WiMAX, WiRAN, HSDPA, HSUPA, Resource allocation and interference management, Quality of service and scheduling methods, Capacity planning and dimensioning, Cross-layer design and Physical layer based issue, Interworking architecture and interoperability, Relay assisted and cooperative communications, Location and provisioning and mobility management, Call admission and flow/congestion control, Performance optimization, Channel capacity modeling and analysis, Middleware Issues: Event-based, publish/subscribe, and message-oriented middleware, Reconfigurable, adaptable, and reflective middleware approaches, Middleware solutions for reliability, fault tolerance, and quality-of-service, Scalability of middleware, Context-aware middleware, Autonomic and self-managing middleware, Evaluation techniques for middleware solutions, Formal methods and tools for designing, verifying, and evaluating, middleware, Software engineering techniques for middleware, Service oriented middleware, Agent-based middleware, Security middleware, Network Applications: Network-based automation, Cloud applications, Ubiquitous and pervasive applications, Collaborative applications, RFID and sensor network applications, Mobile applications, Smart home applications, Infrastructure monitoring and control applications, Remote health monitoring, GPS and location-based applications, Networked vehicles applications, Alert applications, Embeded Computer System, Advanced Control Systems, and Intelligent Control : Advanced control and measurement, computer and microprocessor-based control, signal processing, estimation and identification techniques, application specific IC’s, nonlinear and adaptive control, optimal and robot control, intelligent control, evolutionary computing, and intelligent systems, instrumentation subject to critical conditions, automotive, marine and aero-space control and all other control applications, Intelligent Control System, Wiring/Wireless Sensor, Signal Control System. Sensors, Actuators and Systems Integration : Intelligent sensors and actuators, multisensor fusion, sensor array and multi-channel processing, micro/nano technology, microsensors and microactuators, instrumentation electronics, MEMS and system integration, wireless sensor, Network Sensor, Hybrid

Sensor, Distributed Sensor Networks. Signal and Image Processing : Digital signal processing theory, methods, DSP implementation, speech processing, image and multidimensional signal processing, Image analysis and processing, Image and Multimedia applications, Real-time multimedia signal processing, Computer vision, Emerging signal processing areas, Remote Sensing, Signal processing in education. Industrial Informatics: Industrial applications of neural networks, fuzzy algorithms, Neuro-Fuzzy application, bioInformatics, real-time computer control, real-time information systems, human-machine interfaces, CAD/CAM/CAT/CIM, virtual reality, industrial communications, flexible manufacturing systems, industrial automated process, Data Storage Management, Harddisk control, Supply Chain Management, Logistics applications, Power plant automation, Drives automation. Information Technology, Management of Information System : Management information systems, Information Management, Nursing information management, Information System, Information Technology and their application, Data retrieval, Data Base Management, Decision analysis methods, Information processing, Operations research, E-Business, E-Commerce, E-Government, Computer Business, Security and risk management, Medical imaging, Biotechnology, Bio-Medicine, Computer-based information systems in health care, Changing Access to Patient Information, Healthcare Management Information Technology. Communication/Computer Network, Transportation Application : On-board diagnostics, Active safety systems, Communication systems, Wireless technology, Communication application, Navigation and Guidance, Vision-based applications, Speech interface, Sensor fusion, Networking theory and technologies, Transportation information, Autonomous vehicle, Vehicle application of affective computing, Advance Computing technology and their application : Broadband and intelligent networks, Data Mining, Data fusion, Computational intelligence, Information and data security, Information indexing and retrieval, Information processing, Information systems and applications, Internet applications and performances, Knowledge based systems, Knowledge management, Software Engineering, Decision making, Mobile networks and services, Network management and services, Neural Network, Fuzzy logics, Neuro-Fuzzy, Expert approaches, Innovation Technology and Management : Innovation and product development, Emerging advances in business and its applications, Creativity in Internet management and retailing, B2B and B2C management, Electronic transceiver device for Retail Marketing Industries, Facilities planning and management, Innovative pervasive computing applications, Programming paradigms for pervasive systems, Software evolution and maintenance in pervasive systems, Middleware services and agent technologies, Adaptive, autonomic and context-aware computing, Mobile/Wireless computing systems and services in pervasive computing, Energy-efficient and green pervasive computing, Communication architectures for pervasive computing, Ad hoc networks for pervasive communications, Pervasive opportunistic communications and applications, Enabling technologies for pervasive systems (e.g., wireless BAN, PAN), Positioning and tracking technologies, Sensors and RFID in pervasive systems, Multimodal sensing and context for pervasive applications, Pervasive sensing, perception and semantic interpretation, Smart devices and intelligent environments, Trust, security and privacy issues in pervasive systems, User interfaces and interaction models, Virtual immersive communications, Wearable computers, Standards and interfaces for pervasive computing environments, Social and economic models for pervasive systems, Active and Programmable Networks, Ad Hoc & Sensor Network, Congestion and/or Flow Control, Content Distribution, Grid Networking, High-speed Network Architectures, Internet Services and Applications, Optical Networks, Mobile and Wireless Networks, Network Modeling and Simulation, Multicast, Multimedia Communications, Network Control and Management, Network Protocols, Network Performance, Network Measurement, Peer to Peer and Overlay Networks, Quality of Service and Quality of Experience, Ubiquitous Networks, Crosscutting Themes – Internet Technologies, Infrastructure, Services and Applications; Open Source Tools, Open Models and Architectures; Security, Privacy and Trust; Navigation Systems, Location Based Services; Social Networks and Online Communities; ICT Convergence, Digital Economy and Digital Divide, Neural Networks, Pattern Recognition, Computer Vision, Advanced Computing Architectures and New Programming Models, Visualization and Virtual Reality as Applied to Computational Science, Computer Architecture and Embedded Systems, Technology in Education, Theoretical Computer Science, Computing Ethics, Computing Practices & Applications Authors are invited to submit papers through e-mail [email protected]. Submissions must be original and should not have been published previously or be under consideration for publication while being evaluated by IJCSIS. Before submission authors should carefully read over the journal's Author Guidelines, which are located at http://sites.google.com/site/ijcsis/authors-notes .

Journal of Computer Science December 2013

Documents

Transcript of Journal of Computer Science December 2013