COLOR HISTOGRAM BASED TEXT DETECTION IN IMAGES

96

This article can be downloaded from http://www.ijeetc.com/currentissue.php

Int. J. Elec&Electr.Eng&Telecoms. 2014 Bhagyalaxmi P T and Bhaskar B, 2014

COLOR HISTOGRAM BASED TEXT DETECTION INIMAGES

Bhagyalaxmi P T1* and Bhaskar B1

*Corresponding Author: Bhagyalaxmi P T,[email protected]

Texts in characteristic scene pictures convey rich semantic data, which might be utilized to helpan extensive variety of provisions, for example, object distinguishment, picture recovery, andhuman machine cooperation. Text in many pictures taken commonly by a cell phone can havea critical relative or point of view twisting, making the current content location and the consequentOCR motors inclined to disappointments. Not at all like other content recognition approachesthat for the most part depend on either content structure or color, we propose an edge work thatjoins together both by overseeing content situated shade diminishment process with adjustedSWT data. SWT pixels mapped to shade space vote for the color they compare to. Shadesaccepting high SWT vote in all probability fit in with content zones and are obstructed from beingmean-moved away and than applying composition invariant low rank convert that can identifyand correct messages in self-assertive introductions in the picture against complex foundations,so that the writings could be accurately perceived by basic OCR motors.

Keywords: Natural scene pictures, SWT, Color decrease, Texture invariant low rank transform,Arbitrary introduction

INTRODUCTIONContent Extraction from picture is concernedwith concentrating the pertinent contentinformation from an accumulation of pictures.These days the extent of the accessiblecomputerized feature substance is expandingquickly. This prompts a pressing requirementfor quick and compelling calculations for datarecovery from interactive media content.Printed data in interactive media constitutes

ISSN 2319 – 2518 www.ijeetc.comVol. 3, No. 3, July 2014

© 2014 IJEETC. All Rights Reserved

Int. J. Elec&Electr.Eng&Telecoms. 2014

1 Department of Electronics &Communication Engineering, SJBIT, Bangalore, Karnataka, India.

an exceptionally rich wellspring of abnormalamount semantics for recovery and indexing.Complex foundations, uneven light, and vicinityof just about boundless number of content textstyles, sizes, and introductions postureincredible troubles even to state-of-the-craftsmanship content identif icationtechniques. Not at all like record pictures,where content is typically superimposed oneither clear or complex foundations and is in

Research Paper

97



this manner more unique (Chen and Wu, 2009;and Zagoris et al., 2011), regular scenepictures bargain with scene content, which isas of now a piece of the caught scene and isregularly a great deal less different. All thingsconsidered, content discovery has turned intoan extremely prominent examination territorybecause of its gigantic potential in numerouspractical ranges, for example, signinterpretation, substance based web picturelooking, and helping the outwardly debilitated.State-of-the-craftsmanship writing recognizestwo significant content discovery approaches:composition based and area based. Surfacebased systems (Chen and Yuille, 2004; andYe et al., 2005) sweep pictures at diversescales, review zone under the sliding windowfor content like gimmicks, and arrange it ascontent/non-message. They frequently needaccuracy and are moderately moderatebecause of their scale-space approach.District based strategies (Epshtein et al.,2010; and Yi and Tian, 2011), then again, workin a base up style by selecting pixels (or areas)with ordinary content properties and gatheringthem into joined segments that are furthergeometrically separated and assembled intocontent lines and/or words. Area basedroutines are not constrained to content size/introduction and are (contrasted withcomposition based strategies) quicker. Otherthan the previously stated methodologies,mixture methodologies exist, which adventurepreferences of both composition based andregion based methodologies (Pan et al.,2011). In the vision group, it is for the most partaccepted that content location anddistinguishment have been a broadlyconcentrated on issue and developed results

should as of now exist. This couldn’t have beena long way from reality. For one, most existingcontent recognition and distinguishmentframeworks accept that the writings in thepicture are taken from an about frontalperspective. This is frequently a long way fromthe case for web pictures or pictures taken bycellular telephones, The writings in the picturecould be extremely contorted by a subjectiveplanar (revolution, relative, or homography)convert from the picture plane. Practically allcurrent content recognition routines performinadequately on such pictures. Regardless ofthe possibility that such messages are by onemeans or another discovered the resultingdistinguishment would additionally fall flat inlight of the fact that the writings are not in theirupright position which is frequently required byOCR motors for correct distinguishment. Topurpose this issue, we propose a straightforward yet compelling plan that coordinatesSWT and TILT systems and can naturallyseparate just about writings in a picture notwithstanding their self-assertive beginningmiss hapenings.

METHODOLOGYText in common scene pictures is recognizedfrom other picture structures and foundationby its trademark shape (character strokes arepretty much parallel) and shade consistency.(Epshtein et al., 2010) is an area basedcontent location system. It takes after the strokewidth consistency supposition, which statesthat stroke widths stay consistent all throughindividual content characters. In the wake ofgetting an edge guide of a data picture, SWTstrategy places sets of parallel edge pixels inthe accompanying manner: for each one edgepixel p an inquiry beam in the edge inclination

98



course is genera ted, and the first edge pixelq along the hunt beam is found. In the eventthat p and q have about inverse anglebearings, an edge pair is structured and theseparation in the middle of p and q (calledstroke width) is figured. All pixels lying on theinquiry beam in the middle of p and q (countingp and q) are allocated relating stroke width. Inthe wake of doling out stroke widths to allpicture pixels, the SWT strategy bunchespixels with comparable stroke widths intoassociated segments and channels out thosethat disregard geometrical properties of thecontent. At the point when the edge limit issufficiently low, SWT ordinarily discovers allcharacters in the picture or in any event littlepartitions of each of them.

Text in scene pictures is recognized fromother picture structures and foundation by itstrademark shape (character strokes muchparallel) and color consistency. Epshtein et al.(2010) is an area based content recognitiontechnique.

Text in regular scene pictures is recognizedfrom other picture structures and foundationby its trademark shape (character strokes arepretty much parallel) and color consistency. [8]is an area based content recognitiontechnique. It takes after the stroke widthsteadiness presumption, which states thatstroke widths stay consistent all throughindividual content characters. In the wake ofgetting an edge guide of an info picture, SWTtechnique places sets of parallel edge pixelsin the accompanying manner: for each oneedge pixel p a pursuit beam in the edgeinclination course is genera ted, and the firstedge pixel q along the hunt beam is found. Inthe event that p and q have about inverse angleheadings, an edge pair is framed and theseparation in the middle of p and q (calledstroke width) is After getting the SWT picture,all pixels with non-zero SWT qualities aremapped to the RGB shade space utilizingSWT lookup table Within request not to mean-movement genuine content shades, SWTvoting is performed at every mean shift

Figure 1: Block Diagram of Proposed System

Original Image RGB Histogram Initial ColorCubes

Mean Shift withSWT Voting

Rectification usingTILT

Text Detection

SWT Filter SWT DirectionDetermination

99



emphasis. On the off chance that the sourceand target blocks accept high and low amountsof SWT votes, individually, the shade ispresumably leaving the safe content zone andis hindered from moving any further. In differentcases, mean-moving is permitted. In place forSWT voting to work rightly, it is vital that SWTpicture compares to genuine contentcharacters. To manage both dim and lightcontent situations, the first SWT execution runsthe entire content discovery flowchart twice—in slope and counter-angle bearings—andconsolidates the after effects of both headings.Consequently, we propose a SWT bearingdetermination strategy, which gives right SWTpicture to SWT voting stage. We actualizedthe Nikolaou shade decrease system withchanges. The introductory solid shapes are notchosen haphazardly as in Nikolaou andPapamarkos (2009) however in theaccompanying way: RGB histogramreceptacles are sorted in sliding request andstarting 3d square focuses are constantlychosen from the highest point of the unvisitedcanisters rundown. Later stage we apply tiltcalculation to get back just the content areafrom unique picture in frontal perspective pointalongside shade.

OUTLINE ANDIMPLEMENTATIONFigure 2 shows the procedure stream outlineof the entire framework which starts with theinfo picture. At first the information color pictureis changed over to ash scale then the pictureis at the same time given as info to shadedecrease process and stroke width changeprocess. The yield of both the procedure areconsolidated and dispensing with of noncontent area is carried out focused around swt

voting. Here at this stage color shapes shapedfor a picture just the genuine nature (contentcolors) are held are matched with stoke widthresult and non message areas are meanmoved focused around swt voting. The yieldacquired at this stage is color picture holdingjust content district however the non caughtcontent districts are still present which areredetected utilizing texture invariant low ranktransform.

Start

RGB to Gray Color Conversion

Area Based NoiseRemoval

Initial Color Cubes-PreservingDominant Text Colors

SWT DirectionDetermination

Mean Shift withSWT Voting

Rectification of theDetected Text Using TILT

Stop

RGB Histogram SWT Filter

Figure 2: Proposed System Flowchart

100



EXPERIMENTSIn this section, we present extensiveexperiments to verify the effectiveness of oursystem in extracting arbitrarily oriented texts innatural images, in comparison with many of thestate-of-the-art systems and methods. Inaddition, we verify empirically how the resultsof our system could significantly impact on theperformance of OCR engines. A commonlyadopted way to evaluate algorithms is to usethe f-measure, which is a combination ofprecision and recall. They are defined as follows:

Precession = TP/(TP + FP)

Recall = TP/(TP + FN)

fMeasure = (2 * Precission * Recall)/(Precision + Recall)

where TP is true positive, FP is false positiveand FN is false negative.

This dataset consists of text areas detectedby SWT. SWT mainly detects texts that arenearly horizontal or with slight distortion.

This dataset consists of text areas detectedby method (Chen and Wu, 2009). This methodmainly detects horizontal texts and can alsodetect slight distorted text regions.

This dataset consists of text areas detectedby Alex Chen’s (Chen and Yuille, 2004)method.

Figure 3: Epshtein Dataset

Figure 4: Yi Dataset

Figure 5: Alex Chen Dataset

Figure 6: Unrectified Dataset

Figure 7: Rectified Dataset

101



This dataset consists of texts detected byour method in the original images. Most ofthem are not in upright position.

This dataset consists of the correspondingrectified text outputs from our method. Almostall are correctly rectified to their uprightposition.

Algorithm Precision Recall f-measure

Our System 0.8333 0.9090 0.8695

Xin Zhang 0.7956 0.8299 0.8124

Epshtein 0.5501 0.6164 0.5814

Yi 0.3624 0.5686 0.4427

Alex Chen 0.3658 0.3254 0.3444

Table:1: Performances of Different TextDetection Methods

CONCLUSIONOur framework significantly beat existingcontent discovery frameworks on datasets thathold commonsense, uncontrolled pictures fromthe Internet or taken by cell telephones. Limitsin the SWT voting condition focus affectabilityof the content/non-content move identificationin the mean-movement stage. On the offchance that the edges are loose, considerablymore content colors could be protected,however for this situation content charactersare regularly part into a few content shades.Our future work will subsequently concentrateon uniting content shades that fit in with thesame content character.

REFERENCES1. Chen X and Yuille A L (2004), in

Proceedings of the IEEE Conference onComputer Visionand Pattern Recognition,Detecting a Reading Text in NaturalScenes, pp. 366-373, IEEE, Washington.

2. Chen Y Land Wu B F (2009), “A Multi-PlaneApproach for Text Segmentation ofComplex Document Images”, PatternRecognition, Vol. 42, No. 7, pp. 1419-1444.

3. Epshtein B, Ofek E and Wexler Y (2010a),in IEEE Conference on Computer Visionand Pattern Recognition, Detecting Textin Natural Scenes with Stroke WidthTransform, June 13-18, pp. 2963-2970,IEEE, San Francisco, CA, Washington.

4. Epshtein B, Ofek E and Wexler Y (2010b),“Detecting Text in Natural Scenes withStroke Width Transform”, in CVPR,pp. 2963-2970.

5. Nikolaou N and Papamarkos N (2009),“Color Reduction for Complex DocumentImages”, Int. J. Imaging Syst. Technol.,Vol. 19, No. 1, pp. 14-26.

6. Pan Y F, Hou X and Liu C L (2011), “A HybridApproach to Detect and Localize Texts inNatural Scene”, IEEE Trans. ImageProcess, Vol. 20, No. 3, pp. 800-813.

7. Ye Q, Huang Q, Gao W and Zhao D (2005),“Fast and Robust Text Detection in Imagesand Video Frames”, Image and VisionComput., Vol. 23, No. 6, pp. 565-576.

8. Yi C and Tian Y (2011), “Text StringDetection from Natural Scenes byStructure-Based Partition and Grouping”,IEEE Trans. on Image Process, Vol. 20,No. 9, pp. 2594-2605.

9. Zagoris K, Chatzichristofis S A andPapamarkos N (2011), “Text LocalizationUsing Standard Deviation Analysis ofStructure Elements and Support VectorMachines”, EURASIP J. Adv. SignalProcess, Vol. 47, pp. 1-12.

COLOR HISTOGRAM BASED TEXT DETECTION IN IMAGES

Documents

Transcript of COLOR HISTOGRAM BASED TEXT DETECTION IN IMAGES