Off-Line Signature Recognition UsingMorphological Pixel Variance Analysis

H B Kekre, V A BharadiComputer Science DepartmentMPSTME, NMIMS University

Mumbai, India+91-9323557897, +91-

9819125676hbk@yahoo.com,

vinu_bharadi@rediffmail.com

S GuptaComputer Engineering

DepartmentTCET, Mumbai University

Mumbai, India+91-9820242661

shiwanigupta_om@live.com

A A Ambardekar, V BKulkarni

Electronics DepartmentFAMT, Mumbai University

Ratnagiri, India+91-9422431885

vinay291970@rediffmail.com

ABSTRACTHandwritten signatures are one of the oldestbiometric traits for human authorization andauthentication of documents. Majority ofcommercial application area deal with staticform of signature. In this paper we presenta method for off-line signature recognition.We have used morphological dilation onsignature template for measurement of thepixel variance and hence the inter class andintra class variations in the signature. Theproposed feature extraction mechanism isfast enough so that it can be applied foron-line signature verification also.

Categories and SubjectDescriptors I.4.7 Image Processing and Computer vision

General TermsAlgorithms, Design, Experimentation,Security, Human Factors, Verification.

KeywordsBiometrics, Static Signature Recognition,Morphology.

1. INTRODUCTIONA problem of personal verification andidentification is an actively growing areaof research. The methods are numerous, andare based on different personalcharacteristics. Voice, lip movements, handgeometry, face, odor, gait, iris, retina,fingerprint are the most commonly usedauthentication methods. All of these and

behavioral characteristics are calledbiometrics.

BiometricsThe biometrics is most commonly defined asmeasurable psychological or behaviouralcharacteristic of the individual that can beused in personal identification andverification. The driving force of theprogress in this field is, above all, thegrowing role of the Internet and therequirements of society. Therefore,considerable applications are concentratedin the area of electronic commerce andelectronic banking systems and securityapplications of vital installations.

The biometrics has a significant advantageover traditional authentication techniques(namely passwords, PIN numbers, smartcardsetc.) due to the fact that biometriccharacteristics of the individual are noteasily transferable, are unique of everyperson, and cannot be lost, stolen orbroken. The choice of one of the biometricsolutions depends on several factors [2]:

User acceptance Level of security required Accuracy Cost and implementation time

Biometric and biomedical informatics are thefast developing scientific direction,studying the processes of creation,transmission, reception, storage,processing, displaying and interpretation ofinformation in all the channels offunctional and signal systems of livingobjects which are known to biological andmedical science and practice. Modern naturalsciences at present sharply need in theupdating of scientific picture of the world,and the essential contribution in thisprocess can be made by the biometric andbiomedical methods. Only some more simple(statistical) forms of biometric andbiomedical information have found theirapplication when person identification, andraised interest for these methods ofidentification can be caused by newpossibilities of information technologies.

Handwritten Signature RecognitionHandwritten signature verification has beenextensively studied & implemented. Its manyapplications include banking, credit cardvalidation, security systems etc. Ingeneral, handwritten signature verificationcan be categorized into two kinds – on–lineverification and off–line verification [3][10][21]. On–line verification requires astylus and an electronic tablet connected toa computer to grab dynamic signatureinformation [21]. Off–line verification, onthe other hand, deals with signatureinformation which is in a static format.In On–line approach we can acquire moreinformation about the signature whichincludes the dynamic properties ofsignature. We can extract information aboutthe writing speed, pressure points, strokes,acceleration as well as the staticcharacteristics of signatures [22]. Thisleads to better accuracy because the dynamiccharacteristics are very difficult toimitate, but the system requires user co-

ICWET’10, February 26 - 27, 2010, Mumbai,Maharashtra, India.Copyright 2010 ACM 978-1-60558-812-4…$5.00.

Permission to make digital or hard copies of allor part of this work for personal or classroomuse is granted without fee provided that copiesare not made or distributed for profit orcommercial advantage and that copies bear thisnotice and the full citation on the first page.To copy otherwise, or republish, to post onservers or to redistribute to lists, requiresprior specific permission and/or a fee.

operation and complex hardware. Digitizertablets or pressure sensitive pads are usedto scan signature dynamically, one suchtablet is shown in Figure 1.

Figure 1. Digitizer Tablet for On-lineSignature Scan

In off–line signature recognition we arehaving the signature template coming from animaging device, hence we have only staticcharacteristic of the signatures. The personneed not be present at the time ofverification. Hence off-line signatureverification is convenient in varioussituations like document verification,banking transactions etc. [1][12][13][14].As we have a limited set of features forverification purpose, off-line signaturerecognition systems need to be designed verycarefully to achieve the desired accuracy.

Steps in Signature Recognition [12][22]Signature Recognition Systems need topreprocess the data. It includes a series ofoperations to get the results. The majorsteps are as follows

Data Acquisition The signatures to be processed by thesystem should be in the digital imageformat. We need to scan the signatures fromthe document for the verification purpose Signature Pre-processing We have to normalize the signature, resizeit to proper dimensions, remove thebackground noise, and thin the signature.This yields a signature template which canbe used for extracting the features. Atypical scanned and Pre-Processed Signatureis shown in Figure 2.

Figure 2. Pre-processing of a signature

Feature Extraction In this paper we are using morphologicalcontours of the signature and theiroverlapping map as a feature vector. Thesecontours are generated by multiplemorphological dilation of the signature.

Enrollment & Training The extracted features are stored in todatabase. The human signature is dependenton varying factors, the signaturecharacteristics change with thepsychological or mental condition of aperson, physical and practical conditionlike tip of the pen used for signature,signatures taken at different times, agingetc. We have to consider a high degree of intra-class variation because two signatures froma same person are never same. Our systemshould consider this variation and at thesame time the system should possess highdegree of accuracy to detect forgedsignatures.

Figure 3. Simplified workflow for a typicalSignature Recognition System

We train the system using a training set ofsignature obtained from a person. Designingof a classifier is a separate area ofresearch. The decision thresholds requiredfor the classification are calculated byconsidering the variation of features amongthe training set. Separate set of thresholds(user Specific) is calculated for eachperson enrolled, some system also use commonthreshold for all users.

Performance Evaluation The performance of system depends on howaccurately the system can classify betweenthe genuine and fraud signatures. Theforgeries involved in handwritten signatureshave been categorized based on theircharacteristic features [5].

2. OFF-LINE SIGNATURE RECOGNITIONA lot of research has been done in the fieldof Off-line signature recognition. This is aconvenient approach and various optimizationtechniques are applied to address theproblem. Sabourin [23] used granulometricsize distributions for the definition oflocal shape descriptors in an attempt tocharacterize the amount of signal activityexciting each rectangle on the focus of asuperimposed grid. He then used a nearest

neighbor and threshold-based classifier todetect random forgeries. A total error rateof 0.02% and 1.0% was reported for therespective classifiers. A database of 800genuine signatures from 20 writers is used.Abbas [20] used a back propagation neuralnetwork prototype for the offline signaturerecognition. He used feed forward neuralnetworks and three different trainingalgorithms Vanilla, Enhanced and batch wereused. In [20] he reported FAR between therange of 10-40 % for casual forgeries. Aneuro-fuzzy system was proposed by Hanmandlu[30], they compared the angle made by thesignature pixels that are computed withrespect to reference points and the angledistribution was then clustered with fuzzyc-means algorithm. Back propagationalgorithm was used for training neuralnetwork. The system reported FRR in therange of 5-16% with varying threshold. Zhang [6] have proposed a Kernel PrincipalComponent Self-regression (KPCSR) model foroff-line signature verification andrecognition problems. Developed from theKernel Principal Component Regression(KPCR), the self-regression model selected asubset of the principal components from thekernel space for the input variables toaccurately characterize each person’ssignature, thus offering good verificationand recognition performance. The modeldirectly worked on bitmap images in thepreliminary experiments, showingsatisfactory performance. A modular schemewith subject-specific KPCSR structure provedto be very efficient, from which each personwas assigned an independent KPCSR model forcoding the corresponding visual information.He reported FRR 92% and FAR .5%

Baltzakis [14] developed a neural network-based system for the detection of randomforgeries. The system uses global features,grid features (pixel densities), and texturefeatures (co occurrence matrices) torepresent each signature. For each one ofthese feature sets, a special two-stageperceptron one-class-one-network (OCON)classification structure is implemented. Inthe first stage, the classifier combines thedecision results of the neural networks and

the Euclidean distance obtained using thethree feature sets. The results of the firststage classifier feed a second-stage radialbasis function (RBF) neural networkstructure, which makes the final decision. Adatabase is used which contains thesignatures of 115 writers, with between 15and 20 genuine signatures per writer. Anaverage FRR and FAR of 3%and 9.8%,respectively is obtained. In [25] Armand,Blumenstein and Muthukkumarasamy usedcombination of the Modified DirectionFeature (MDF) in conjunction with additionaldistinguishing features to train and testtwo Neural Network-based classifiers. AResilient Back Propagation neural networkand a Radial Basis Function neural networkwere compared. Using a publicly availabledatabase of 2106 signatures containing 936genuine and 1170 forgeries, they obtained averification rate of 91.12%.Justino [15] used a discrete observation HMMto detect random, casual, and skilledforgeries. A grid segmentation scheme wasused to extract three features: a pixeldensity feature, a pixel distributionfeature (extended-shadow-code), and an axialslant feature. A cross-validation procedurewas used to dynamically define the optimalnumber of states for each model (writer).Two data sets are used. The first data setcontains the signatures of 40 writers with40 genuine signatures per writer. This dataset was used to determine the optimalcodebook size for detecting randomforgeries. This optimized system was thenused to detect random, casual, and skilledforgeries in a second data set. The seconddata set contains the signatures of 60writers with 40 training signatures, 10genuine test signatures, 10 casualforgeries, and 10 skilled forgeries perwriter. An FRR of 2.83%and an FAR of 1.44%,2.50%, and 22.67% are reported for random,casual, and skilled forgeries, respectively.Kaewkongka, Chamnongthai and Thipakom [30]proposed a method of off-line signaturerecognition by using Hough transform todetect stroke lines from signature image.The Hough transform was used to extract theparameterized Hough space from signature

skeleton as unique characteristic feature ofsignatures. In the experiment, the BackPropagation trained Neural Network was usedas a tool to evaluate the performance of theproposed method. The system was tested with70 test signatures from different persons.The experimental results reveal therecognition rate 95.24%

Fang [7] developed a system that is based onthe assumption that the cursive segments offorged signatures are generally less smooththan that of genuine ones. Two approachesare proposed to extract the smoothnessfeature: a crossing method and a fractaldimension method. The smoothness feature isthen combined with global shape features.Verification is based on a minimum distanceclassifier. An iterative leave-one-outmethod is used for training and for testinggenuine test signatures. A database with 55writers is used with 24 training signaturesand 24 skilled forgeries per writer. An AERof 17.3% is obtained. Ferrer, Alonso, and Travieso [16], usedOffline Geometric Parameters for AutomaticSignature Verification Using Fixed-PointArithmetic. They used set of geometricsignature features for offline automaticsignature verification based on thedescription of the signature envelope andthe interior stroke distribution in polarand Cartesian coordinates. The feature setwas calculated using 16 bits fixed-pointarithmetic and tested with differentclassifiers, such as hidden Markov models,support vector machines, and Euclideandistance classifier. FRR reported was 2.12%and FAR was 3.13%. S. Audet, P. Bansal, andS. Baskaran [40], designed Off-LineSignature Verification and Recognition usingSupport Vector Machine. They used global,directional and grid features of signatures.Support Vector Machine (SVM) was used toverify and classify the signatures and aclassification ratio of 0.95 was obtained.Deng [19] developed a system that used aclosed contour tracing algorithm torepresent the edges of each signature withseveral closed contours. The curvature dataof the traced closed contours weredecomposed into multi-resolution signals

using wavelet transforms. The zero crossingscorresponding to the curvature data wereextracted as features for matching. Astatistical measurement was devised todecide systematically which closed contoursand their associated frequency data weremost stable and discriminating. Based onthese data, the optimal threshold valuewhich controls the accuracy of the featureextraction process was calculated. Matchingwas done through dynamic time warping.Experiments were conducted independently ontwo data sets, one consisting of Englishsignatures and the other consisting ofChinese signatures. For each experiment,twenty-five writers are used with tentraining signatures, ten genuine testsignatures, ten skilled forgeries, and tencasual forgeries per writer. When only theskilled forgeries are considered, AERs of13.4% and 9.8% are reported for therespective data sets. When only the casualforgeries are considered, AERs of 2.8% and3.0% are reported.Majhi, Reddy and Prasanna [5] proposed amorphological parameter for signaturerecognition, they proposed center of mass ofsignature segments, and the signature wassplit again and again at its center of massto obtain a series of points in horizontalas well as vertical mode. The point sequenceis then used as discriminating feature; thethresholds were selected separately for eachperson. They achieved FRR 14.58% and FAR2.08%. This concept of geometric centers isused in this project, here we extend theconcept to find successive geometric centersof depth 2 and use them as a set of globalfeatures.

Table.1 Performance Comparison with Off LineSignature Recognition Systems

Sr. Approach FAR FRR Accu

racy

1

Signature Recognition usingMorphological Pixel Variance Analysis (Proposed System)

6.56 3.7 94.94

2 Contour Method [28]

11.60

13.20

86.90

3Exterior Contoursand Shape Features[27]

06.90

06.50

93.80

4

Local Granulometric Size Distributions [23]

07.00

05.00 -

5Back-Propagation Neural Network Prototype [21]

10.00

06.00 -

6 Geometric Centers[5]

09.00

14.58 -

7Two-stage neural network classifier [11]

03.00

09.81

80.81

8 Distance Statistics [18]

34.91

28.30

93.33

9Modified Direction Feature[25]

- - 91.12

10Hidden Markov Model and Cross-Validation [15]

11.70

00.64 -

11Discrete Random Transform and a HMM [12]

10.00

20.00 -

12Kernel Principal Component Self-regression [6]

03.40

08.90 -

13Parameterized Hough Transform [30]

- - 95.24

14Smoothness Index Based Approach [8]

- - 79.00

15Geometric based on Fixed-Point Arithmetic [16]

4.9-15.5

5.61-

16.39

-

16HMM and Graphometric Features [13]

23.00

01.00 -

17 Virtual Support Vector Machine

13.0 16.0 -

[26] 0 0 18 Wavelet–based Verification [19]

10.98

05.60 -

19 Genetic Algorithm[31]

01.80

08.51

86.00

Kekre and Pinge used template matchingapproach in [29]. The signature wassegmented in predefined shape templates, inall 40 different templates were consideredfor feature extraction. They used neuralnetwork classifier. Two separate algorithmswere used first algorithm used 40 shapesassociated with each signature, neuralnetwork with 40 input nodes , 25 nodes inhidden layer and 10 nodes in output layerwas used. The other algorithm used ratiovectors for all the signatures and all thesevectors were used to train a neural networkwith 450 input nodes, 230 nodes in hiddenlayer and 10 nodes in output layer. Total 10users database was used for testing, theiralgorithm 1 reported FAR 20% and algorithm 2reported FAR 0%.All of these efforts were towards automatingthe process of handwritten signaturerecognition. We have defined our projectscope previously. Here we try to develop asignature verification system over theguidelines set by these people. Table Igives the summary of all systems performancemetrics.

3. CONTOUR GENERATIONThe scanned signature is first pre-processedto get a normalized binary template. Wefollow a series of operations like noiseremoval using filtering, Scaling,Smoothening, Intensity normalization,Thinning [1][6]. This gives us a binarysignature template. We use the normalizedtemplate for further processing. Thesignature has intra-class variations i.e.Signatures of a same user have variations,but these variations are limited. Forgedsignatures (Simple forgeries) and differentuser’s signatures have vast variations(Inter – Class variations). We try to detectthese variations in signature segments. Thesignature pixels are having specific limitof variation for the genuine set ofsignature and the forgeries have variations

greater than the limit of a specificperson’s signature. We try to classify thesignature based on this variation. Toquantify this variance we have used theproposed morphological [9] technique. We generate a contour of signature; thiscontour is actually the external boundary ofthe signature. Dilation algorithm is usedfor this and various levels of dilations areused. We use morphological dilation processwith three different structuring elements;the structuring elements [9] are circleswith radius r1, r2, r3, r4 these radiicorrespond to the allowed pixel variance.This process gives four dilated sets D1, D2,D3, and D4 as shown below,

(1)Where S=Signature template Bi = structuringelement- Circle of Radius ri. i=1, 2, 3.This is achieved in programming environmentby drawing circles of various radii on thetemplates and filling them with appropriatecolour; the circles are drawn with radii r1,r2, r3, r4. Where r4 > r3 > r2 > r1. Thisoperation gives a structure with bands ofvarying thickness. These bands of colourswill represent the variation extent of eachpixel and hence the signature segments. Thisstructure is shown in Fig. 4.The four bandsin the testing program were generated withr1=3 Pixel, r2=6 Pixel, r3=10 Pixel andr4=16 Pixel radius and filled with Black,Red, Green, Blue Colours Respectively.

Figure 4. Check Pattern for standardSignature shown in Figure 2.

The colours are represented in (R, G, B)format where Red =(255,0,0) ,Green=(0,255,0), Blue=(0, 0, 255) andwhite=(255,255,255), Black=(0, 0, 0). Wecall this bands pattern as a “Check

pattern”. In the next section we discussdetection process.

Pixel Variance AnalysisFor training purpose we take three standardsignatures from a user and generate dilationsets Di and same band structure for them.For detection purpose we use two templatesat a time, one is the test signature and onefrom the standard templates. Similar checkpattern is generated for the test signature.The test template ‘T’ will be used togenerate the dilation sets we get four testdilation sets as discussed above with thehelp of structuring elements

(2)TDi= test dilation pattern, Bi = structuringelement- Circle of Radius ri. i=1, 2, 3, 4.Now consider each of this dilation set as anallowable pixel variation area, we have thelowest dilation band D1 as the 3 pixeldilated signature template as the allowedvariation and if any pixel is going outsidethis region will be a misplaced pixel ofhigh variation pixel. If the pixel from testsignature are going in the higher dilationbands of original signature Di then theyhave variation more than the specifiedradius of the band, we can count the numberof pixels from test signature which arehaving such variation by set intersectionprocess. We take the intersection ofdilation sets and try to find the crossoccurrence. This is given by

…….i,j=1, 2 ,3, 4. (3)The lowest variation band is C00 , We findthe element count for each set for Cij ifi≠j then it is cross intersection andcorresponding to the variation pixel. Thevalues of i, j correspond to differentdegree of variation. We count the number ofpixel in each band.For genuine signatures with least variationthe element count of Cij where i=j is veryhigh and C00 contain the perfect of leastvariation pixels. Higher is the element

count higher is the matching. If the elementcontent of cross intersection is high i.e.Cij , where i≠j then the pixels are havingmore variation and correspond to mismatch.Each set is given a weight Wij correspondingto degree of match, we evaluate the matchingscore S as follows, Wij (i=j) > Wij (i≠j).

(4)Values of Wij are based on intra classvariation and evaluated by correlation ofsame person’s signature and evaluating thevariation or we can set static values bytrial & error process. The higher the valueof ‘S’ higher is the matching between thetemplate. We can set a threshold for ‘S’ toclassify the signatures.In Table 5, we have shown one personssignature taken for calculation of the abovediscussed parameters, the limits for theacceptance levels are Limit 1 to Limit 5,depend on the intra class variations. Thefeature extraction process is discussed innext section.

Practical ImplementationWe have implemented the algorithm mentionedabove in Visual Basic 6.0. To find theintersection of sets we evaluate the bitwiseExclusive OR (EX-OR) of the colour bandsgenerated by the dilation process. This isillustrated as follows, First templatesign1(x, y) …..x=0,1…..N , y=0,1…..M(5)For a M Rows X N Columns template. Second template sign2(x, y) …..x=0,1…..N , y=0,1…..M(6) for a M rows X N Columns template.EX-OR operation will generate third templatetest(x, y) where

(7)

This EX-OR operation will be performed onthe (R, G, B) colour triplet of each pixelbecause of this in the test(x, y) templatevarious colours are generated by the EX-ORing of R G B bands R(255,0,0) Ex-OrG(0,255,0) will yield (255,255,0) If samecolour is there it will generate White(0, 0,0).Black (0, 0, 0) Ex-Ored with Any of the R,G, B bands will give the same colour onlyi.e. R, G, B only. The R G B bands in signtemplate are actually the allowed variationsfor the pixel and if the pixel is in theallowed variations it will generate Eitherof the pure R, G, B or otherwise anycombination of R, G, B. Pure Black pixels are un-deviated pixels(Black (0, 0, 0) Ex-Or Black (0, 0, 0) Willgive (0, 0, 0) i.e. Black only). By scanningthe check pattern generated we can find outthe pixel variance and hence the matching ofsignature template, The colour codes usedare as follows in Table 2.

Table 2. Colour codes used in programColour R G B0. Black 0 0 01. Red 255 0 02. Green 0 255 03. Blue 0 0 2554. Backgroundcolour 0 100 965. White 255 255 2556. Colour1 0 252 2557. Colour2 255 8 2558. Colour 3 255 252 09. Test resultbackground 255 156 168

4. ResultsWe consider a set of six signatures for aperson as shown in Table 3, Signature 1 to 3are genuine and Signature 4 to 6 areforgery. We perform the dilation based contouranalysis and the check pattern generated isshown in the same figure. Next step of theoperation is to scan this check pattern and

count the number of pixels of each colour, Red (NR), Green (NG), Blue (NB), Black (NBK)and White (NW) respectively.

Table 3. Test Signatures & Corresponding Matching pattern GeneratedSr.No.

Signature Check Pattern Matching Pattern Result

1 Acceptable60.3%

2 Okay64.6%

3 Okay65.32%

4 No Check patternGenerated

RejectedDrasticScaleChange

5 Rejected28.9%

6 Rejected23.29%

Table 4. Results

Parameter Sign1 Sign2 Sign3 Sign4 Sign5 Sign6Black Pixel 43.91 44.37 57.2 0 30.35 29.09Red Pixel 21.78 28.66 30.1 0 24.68 21.01Green Pixel 17.73 25.07 17.68 0 28.84 31.56Blue pixel 9.32 11.51 7.5 0 21.11 27.6Missing orExtra -1.65 -6.04 3.05 0 -13.28 -16.85

OriginalPixels 10.496 10.49 12.26 0 10.137 10.137Test SignPixels 10.661 11.11 11.901 0 11.465 11.795Matching 60.3 64.6 65.32 0 28.96 23.29Remarks Acceptable Okay Okay Rejected Rejected Rejected

Figure 5. Matching Results for Signature 1 to 6

We give these parameters as inputs to a fuzzyclassifier. The fuzzy logic detector has fourparticipation sets Perfect, Good, Acceptableand Rejected, for which the output of fuzzynetwork is given .The matching pattern andthe percentage matching obtained are shown inTable 4 & Figure 5. Normalized parameters areshown in these illustrations. Figure 5 Shows

that Signature 1 to 3 have high degree ofmatching and they are accepted as a match,whereas signatures 4 to 6 are rejected due tolow matching score.

Figure 6. FAR-FRR Plot for SignatureRecognition System (EER = 6.1%)

This method was tested on signatures from thedatabase collected from approx100individuals, from each person minimum 4 tomax 12 signatures were collected, threesignatures each were used for featureextraction & training and remaining signatureused for testing. Forged signatures fromvolunteers were collected for testingpurpose; the database consists of around 1000genuine signatures, 350 different levels offorgeries. Total 257 tests were performed,Table 5 gives the summary.Figure 6 shows theFAR-FRR plot for the system. The system hasachieved EER of 6.1% at 63% Threshold level.Next we present the classified results inTable 6, we present system performancecalculated separately for the Genuine andforged signatures. In the forged signaturegroup we further have group of Casual andSkilled forgery signatures.

Table 5. Signature recognition results

The system is having 100% Accuracy for therejection of casual forgery and for skilledforgery the False Acceptance ratio is 5.79%.For the genuine signatures True AcceptanceRatio is 92.77%.Table 6. Performance Metrics for Final System

Test Samples Ratio

Resultsobtainedon thegiventest bed(%)

Allsampleof asubject

GenuineTAR 92.77FRR 7.23

Forged

Casual

FAR 00.00TRR 100.00

Skilled

FAR 05.79TRR 94.21

5. ConclusionHere we have presented an off-line signaturerecognition system based on the pixelvariance analysis by multiple morphologicaldilations. The system is based on an EX-ORtemplate matching based fuzzy classifier.This is a contour based signature recognitiontechnique. Along with various parameters likenumber of pixels, Angle of rotation, width,height we are using check pattern generatedby multiple morphological dilation. We areusing check pattern to find out variation insignature pixels. The system developed hasreported 94.94% accuracy. The FAR is 6.56%and FRR is 3.7%. The system performance canbe improved further by using more trainingsignature and neuro-computing approach.

6. REFERENCES[1] Abdullah I. Al-Shoshan, “Handwritten

Signature Verification Using ImageInvariants and Dynamic Features”,International Conference on ComputerGraphics, Imaging and Visualization(CGIV'06) ,March 2006

[2] A. K. Jain, A. Ross, S. Prabhakar, “AnIntroduction to Biometric Recognition”,IEEE Transactions on Circuits and Systemsfor Video Technology, Vol. 14, No. 1,January 2004

[3] A. K. Jain, A. Ross, “On Line SignatureVerification”, Pattern Recognition, vol.35, no. 12, pp. 2963-2972, ( 2002).

[4] A. Zimmer and L.L. Ling, "A Hybrid On/OffLine Handwritten Signature VerificationSystem", Seventh International Conferenceon Document Analysis and Recognition,vol.1, pp.424-428, Aug.2003.

[5] B. Majhi, Y. Reddy, D. Babu, “NovelFeatures for Off-line SignatureVerification”, International Journal ofComputers, Communications & Control Vol.I(2006), No. 1, pp. 17-24.

[6] Bai-ling Zhang, “Off-line SignatureRecognition and Verification by KPCSR”,5th International Conference on MachineLearning and Applications (ICMLA'06), pp4 - 6

[7] B. Fang, C.H. Leung, Y.Y. Tang, P.C.K.Kwok, K.W. Tse, Y.K. Wong, “Offlinesignature verification with generatedtraining samples”, IEEE Proc, ImageSignal Process. 149 (2) (2002) 85-90.

[8] B. Fang, Y.Y. Wang, “A Smoothness IndexBased Approach for Off-line SignatureVerification”, Int. Conf. DocumentAnalysis and Recognition, 1999, ICDAR'99,pp 785 - 787,

[9] Jayraman S, Esakkirajan S, veerakumar T,“Digital Image Processing”,tata McGrawHill Publication, 2009, ISBN – 978-0-07-014479-8

RecognitionMode InputsTestSignatures

Accepted/RejectedSignatures

Performance Metrics%

CasesThatShouldbeAccepted

135

CasesActuallyAccepted

130 TAR 96.30

CasesFalselyRejected

05 FRR 3.70

CasesThatShouldbeRejected

122

CasesActuallyRejected

114 TRR 93.44

CasesFalselyAccepted

08 FAR 6.56

[10] H B Kekre, V A Bharadi, “SignatureRecognition using Cluster Based GlobalFeatures”, IEEE International Conference(IACC 2009)

[11] H. Baltzakis, N. Papamarkos, “A newsignature verification technique based ona two-stage neural network classifier”,Engineering Applications of ArtificialIntelligence (2001)

[12] J. Coetzer, B. M. Herbst, J. A. duPreez, “Offline Signature VerificationUsing the Discrete Radom Transform and aHidden Markov Model” , EURASIP Journal onApplied Signal Processing ( 2004),pp 559–571

[13] J Edson, R. Justino, A. El Yacoubi, F.Bortolozzi and R. Sabourin, "An off-lineSignature Verification SystemUsing HMMand Graphometric features", DAS 2000, pp.211-222,

[14] J Edson, R. Justino, F. Bortolozzi andR. Sabourin, "An off-line signatureverification using HMM for Random, Simpleand Skilled Forgeries", SixthInternational Conference on DocumentAnalysis and Recognition (2001) , pp.1031-1034

[15] J Edson, R. Justino, F. Bortolozzi andR. Sabourin, "The Interpersonal andIntrapersonal Variability Influences onOff-line Signature Verification UsingHMM", Proc. XV Brazilian Symp. ComputerGraphics and Image Processing, 2002, pp.197-202,Oct.2002.

[16] Miguel A. Ferrer, Jesu´s, B. Alonso,and Carlos M. Travieso, “OfflineGeometric Parameters for AutomaticSignature Verification Using Fixed-PointArithmetic”, IEEE transactions on patternanalysis and machine intelligence, vol.27, no. 6, June 2005

[17] M. Hanmandlua, Mohd. Hafizuddin, V.Madasu, “Off-line signature verificationand forgery detection using fuzzymodelling” Pattern Recognition 38 (2005),pp 341-356

[18] M.K kalera, S. Shrihari, “OfflineSignature Verification And IdentificationUsing Distance Statistics”, International

Journal of Pattern Recognition and AIVol. 18, No. 7 (2004),pp 1339-1360

[19] P. Deng, H. Yuan Mark Liao & H. Tyan,“Wavelet Based Off-line SignatureRecognition System”, 5th Conference onOptical Character Recognition andDocument Analysis, 1996

[20] R. Abbas, “ BPNN Prototype for offline signature verification”, thesisSubmitted to RMIT, 2003

[21] R. Plamondon, G. Lorette, ‘‘AutomaticSignature Verification and WriterIdentification – The State of the Art”,Pattern Recognition, vol. 4, no. 2, pp.107–131, 1989.

[22] R. Plamondon, “The design of an On-line signature verification system”,International journal of PatternRecognition and Artificial Intelligence,(1994), pp795-811.

[23] R. Sabourin, G. Genest, “Off-linesignature verification localgranulometric size distributions”, IEEETrans. Pattern Anal. Mach. Intell. Vol.19 (1997) pp 976-988.

[24] R.W. Zhou, C. Quek, “Automatic fuzzyneural network driven signatureverification system”, Proceedings of theIEEE Int. Conf. on Neural Networks, 1996,pp. 1034-1039.

[25] S. Armand, M. Blumenstein , “Off-lineSignature Verification based on theModified Direction Feature”, EngineeringApplications of Artificial Intelligence(2004)

[26] S. Audet, P. Bansal, and S.Baskaran ,“Off-line signatureverification using virtual support vectormachines”, ECSE 526 - ArtificialIntelligence, April 7, 2006

[27] S. Chen and S. Srihari, “Use ofExterior Contours and Shape Features inOff-line Signature Verification”, EightInternational Conference on DocumentAnalysis and Recognition (ICDAR’05)

[28] S. Lee and J. C. Pan, ‘‘Off–linetracing and representation ofSignatures,” IEEE Trans. Syst., Man and

Cybern. vol. 22, no. 4, pp. 755–771,July/August 1992.

[29] S. Pinge, H. Kekre, “ SignatureIdentification using Neural Networks”,Proceedings of National Conference onImage Processing 2005 (NCIP2005), Mumbai,India ,pp 31-39

[30] T. Kaewkongka, K. Chamnongthai, “Off-Line Signature Recognition usingparameterized Hough Transform ”,Symposium on Signal Processing and itsApplications, ISSPA ’99

[31] Y. Xuhua, F. Takashi ,“Constructing aHigh Performance Signature VerificationSystem Using a GA Method”, IEEE Conf.ANNES, 20-23 Nov. 1995, PP: 170 - 173