/ IHIIZ!IOH /_OII!HH OI IS!OIIC /IDICPHHSCIIQ! H_CS HSIH_ DCH!IOSOQHIC /QQIOCHhuid N. /min1

, NOhumCd /bd Luuuh?

, /bOu Lu u88uniCn?

und LCiud 5ChuCCi+1 FacultyofComputersandInformation,MenofyaUniversity,Egypt 2FacultyofComputersandInformation,MansouraUniversity,Egypt 3FacultyofComputersandInformation,CairoUniversity,Egypt 4DepartmentofComputerScience,LoughboroughUniversity,U.K. 5ScientifcResearchGroupinEgypt(SRGE),http://www.egyptscience.net Abstrct-Inthispaper,animprovedthresholdingapproach basedonneutrosophicsets(NSs)andadaptivethresholding isproposed.Thisisappliedtodegradedhistoricaldocuments imaginganditsperformanceevaluated.TheinputRGBimage istransformedintotheNSdomain,whichisdescribedusing threesubsets,namelythepercentageoftruthinasubset,the percentageofindeterminacyinasubset,andthepercentageof falsityinasubset.TheentropyinNSisemployedtoevaluate theindeterminacywithaA-meanoperationusedtominimize indeterminacy.Finally,thehistoricaldocumentimageisbinarizedusinganadaptivethresholdingtechnique.Experimental resultsdemonstratethattheproposedapproachisabletoselect appropriateimagethresholdsautomaticallyandeffectively,while itisshowntobelesssensitivetonoiseandtoperformbetter comparedwithotherbinarizationalgorithms. Keywords:imagebinarization,thresholding,historicalmanuscript image,neutrosophictheory. I.INTRODUCTION AncientArabicdocumentstypicallysufferfromvarious degradations due to both ageing anduncontrolledenvironmentalconditions[1].Themainartefactsencounteredindigitally captured images ofhistoricaldocuments are[ l],[2]:shadows, non-uniformillumination,smear,strain,bleed-throughand faintchaacters.Fig.Iillustratessomeexamplesofdegraded historicalArabicmanuscriptimages. Thesedegradationsarise eitherduetothephysicalstorageconditionsoftheoriginal manuscript,orbecauseofwritershavingusedditerentquantitiesofinkandpressureresultingincharacterswithditerent intensitiesand/orthicknessesaswellasfaintcharacters[1]. Inaddition,somedocumentscontainextradetailssuchas diacritics,decorations,orhavewritinginmultiplecolors. Binarizationofsuchdocumentimagesistypicallyanessentialpre-processing task,and henceimportant fordocument analysissystems.Itconvertsanimageintobi-Ievelformin suchwaythattheforeground (text)information isrepresented byblackpixelsandthebackgroundbywhiteones[3]. Althoughdocumentimagebinarizationhasbeenstudiedfor many yeas,thresholding of historical document images isstill a challenging problem due to the complexity of the images and theabovementioneddegradations[2]. Neutrosophicset(NS)approachesarerelativelynewand havebeenappliedtovariousimageprocessingtaskssuchas thresholding,segmentationanddenoising[4].Inthispaper, anewhybridalgorithmforbinarizationofdegradedArabic ,),).- "'VI --'~,-:.'|0)'!Fig.l.Examplesofmanuscriptimagescontainingmulti-coloredtextlines withdifferentdegradations. manuscriptimageisproposed,whichmodifestheprevious algorithmof[5]toworkinanadaptive manner.Experimental results demonstrate that the proposed approach is able to select appropriateimagethresholdsautomaticallyandefectively, whileitisshowntobelesssensitivetonoiseandtoperforms bettercomparedwithotherbinarizationalgorithms. Theremainderofthepaperisstructuredasfollows:SectionsIIandIIIpresentrelatedworkonimagebinarization and neutrosophic sets.InSection IV,theproposed method for binarizationofhistoricaldocumentimagesispresented.Section Vgivesexperimentalresults,whileSection VIconcludes thepaper. II.DOCUMENTIMAGEBINARIZATION Approachesfordocumentimagebinarizationcanbe grouped into two main categories:global and local approaches. Global algorithms selecta single thresholdvalue forthe entire image.Thisgivesgoodresultsifthereisagoodseparation betweenforegroundandbackground.However,forhistorical documents,thisapproachisnotrobustenough[3].Todeal 97-1-4799-94-/14/531.00 Z014 IEEEZwith degradations,thecurrent trend isto use localinformation that guides the threshold value,often pixel-wise, in an adaptive manner[6],[7].Hybrid methods[3] combine globalandlocal informationtoassignpixelstooneofthetwoclasses(textor background). Otsu'sthresholdingalgorithm[8]isthemostwidelyemployedglobalmethod.Ittriestofndthethresholdtwhich separates the gray-level histogram,in an optimal way,into two segments. Itmaximizes the inter-class variance andminimizes theintra-classvariancewiththecalculationofinter-classand intra-classvariancesbasedonthenormalizedhistogramof theimageH = [h...h255]whereL h(i)= 1.Theinter-class varianceisgivenby where and with and (;nter =ql(t)x q2(t)x [/l (t)-/2(t)]2, 1./.( t) =-(-) _h(i)xi, ).t.1 255 /2( t)=-(-) _h(i)xi, q2 t ..).=_

h(i), 255 q2( t)=_h(i). (1 ) (2) (3) (4) (5) Niblack's algorithm [9]calculates a pixelwise threshold in a slidingwindow fashion.Thethresholdt iscomputed byusing themean /andstandarddeviation(ofallthepixelsinthe window,andisderivedas t = / + kx (,(6) wherekisaconstantin[0; 1]thatdetermineshowmuchof thetotalprintobjectedgeisretained. Sauvola'salgorithm[10]isamodifcationofNiblack's approach claimed to give improved performance on documents where the background contains light texture, big variations and unevenillumination.Athresholdiscomputedbasedonthe dynamicrangeRofthestandarddeviationas (7) wherekisafxedvalue.Accordingto[11],thismethodis showntobemoreefectivethanNiblack'salgorithmwhen thegray-levelofthetextiscloseto0andthatofbackground closeto255.However,inimageswherethegray-levelsof backgroundandtextpixelsareclose,theresultsareunsatisfactory. III.NEUTROSOPHICTHEORY Neutrosophy theory considers an event, concept, or entityA in relation to its opposite Anti-A and neutrality Neut-A,which isneither Anor Anti-A.Neut-Aand Anti-Aarereferred to as Non-A.EveryideaAtendstobeneutralizedandbalancedby Anti-AandNon-A.Forexample,ifA= "white",thenAntiA= "black",Non-A= "blue,yellow,red,black,etc."(any colorexceptwhite),andNeut-A= "blue,yellow,red,etc." (anycolorexceptwhiteandblack)[12],[13]. A.NeutrsophicSet LetUbeauniverseofdiscourse,andaneutrosophicset AincludedinU.Anelementxintheset!isdenoted asx( T, I, F).T,IandFarerealstandardornon-standard subsetsof]- 0,1 +[withsup T= Csup,infT=t_inf,sup I= i_sup,infI= i_inf,sup F= f_sup, infF= Linf,n_sup= Csup+Lsup+f_sup,and n_inf= Cinf+i_inf+ f_infoT,IandFarereferredto asneutrosophiccomponents. x( T, I, F)belongstoAinthefollowingway:itist% true in the set, i% indeterminate, andf% false,where t varies in T, ivariesinI,andf variesinF.T,I andF aresubsets,while T,IandFarefunctions/operatorsdependingonknownor unknown parameters.T,I andF arenot necessarilyintervals, andmaybediscreteorcontinuous,single-element,fnite,or countable infnite,union or intersection of various subsets,etc. Theymayalsooverlap. B.NeutrsophicImage LetU be a universe of discourse, andW be a set included in U,whichiscomposed of bright pixels.Aneutrosophic image PNsischaracterizedbythreesubsetT,IandF.ApixelP intheimageisdescribedasP( T, I, F)andbelongstoW in thefollowingway:itist%trueinthebrightpixelset,i% indeterminate,andf%false.ThepixelP( i,j)intheimage domainistransformed intoneutrosophicdomainby PNs(i,j)= {T(i,j),I(i,j), F(i,j)},(8) whereT(i,j),I(i,j)andF(i,j)aretheprobabilitiesof belongingtothebright,indeterminateandnon-brightset, respectively,whicharedefnedas and with T('.) _g(i,j) - gmi n 1J-, gmax- gmi n F(i,j)= 1 -T(i,j), I(i,j)= 1 _ Ho(i,j) - Homin, H Omax-H Omin Ho(i,j) = le(i,j)l, (9) (10) (11) (12) whereH o(i, j) isthehomogeneityvalueofT at(i, j),which isdescribedbythelocalgradientvaluee(i,j),g(i,j)isthe intensityvalueofthepixeI(i,j),gminandgmax arethe minimumandmaximumvalueofg( i, j) respectively. C.NeutrsophicImageEntrpy Th entropy is utilized to evaluate the distribution of diferent graylevelsinanimage.Iftheentropyismaximal,the diferentintensitieshaveequalprobabilityandtheintensities thusdistributeuniformly.Ontheotherhand,iftheentropy issmall,theintensitieshavediferentprobabilitiesandtheir distributionsarenon-uniform. Z7Neutrosophic imageentropy isdefned asthesummation of theentropiesofthethreesubsetT,I andF,andisemployed toevaluatethedistributionoftheelementsintheNSdomain: with and EnT= LPT(i)lnPT(i), Enr = L Pr(i) InPr(i), (13) (14) (15) (16) whereEnT,EnrandEnFaretheentropiesofsubsetsT, IandF,respectively,andPT(i),Pr(i)andPF(i)arethe probabilities of element i in T, I and F, respectively. EnT and EnFareutilizedtomeasurethedistributionoftheelements intheneutrosophicset,andEnrisemployedtoevaluatethe indeterminationdistribution. D.A-meanOpertion Asmentioned,thevalueofI(i,j)isemployedtomeasure theindeterminatedegreeofPNs(i,j).TomakeItobe correlatedwithTandF,changesinTandFinfuencethe distributionofelementsinI andtheentropyofI.Inthegray leveldomain,aA-meanoperation for imageX canbedefned as _1i+w/2 XCi,j)= - L wxw j+w/2 L m=i-w/2n=j-w/2 where wisthelocalwindowsize. X(r, n), AA-meanoperationforPNSisdefnedas: (17) convertedtograylevelimagesPGusingtheNTSCstandard method. Marlme In RGBtoGr c. to theNl sta Fig.2.Overviewoftheproposeddocumentanalysisapproach. A.Pre-prcessing Since historical document images are usually of low quality, apre-processingstageisessentialinordertoeliminatenoise areas,smooththebackgroundtextureandbetterhighlightthe contrastbetweenbackground and textareas[14].Theuseofa low-pass Wiener flter has proved effcient in thiscontext [15], withthewindowsizeoftheflterselectedaccordingtothe minimum character line width [14]. In our method, the window sizeisselectedas3 x3.TheflteredgraylevelimagePgw canthebebinarizedinthenextstage. PNS(A)= P(T(A), I(A), F(A)), (18) B.Binarization with and _ 1Hw/2 TA(i,j) =-L wxw j+w/2 L m=i-w/2 n=j-w/2 _1Hw/2 FA(i,j) =-L wxw j+w/2 L m=i-w/2 n=j-w/2 T(r, n), F(r,n), -( . .) __ Ho(i,j) - HOmin fA Z,]-1 , H 0max-H 0min (19) (20) 21) whereHo(i,j)isthehomogeneityvalueof T(A)at(i,j). AfterthetruesubsetTishandledusingtheA-mean operation,noiseinTisremovedandTbecomesmore homogeneous,andconsequentlymoresuitabletosegmentT preciselyevenusingasimplethresholdingmethod. IV.PROPOSEDApPROACH Fig.2summarisesthevariousstepsoftheproposedalgorithm.CapturedmanuscriptimagesRGBimagesPRGB (ofdifferentsizesandstoredinsimplebitmapformat)are Thisisthestagewhereweemployamodifedversion oftheneutrosophicthresholdingmethodof[5].Thefltered grayimagePgw istransformedintotheneutrosophicdomain, givingPNsasdescribedinSectionIII-B.Then,theA-mean operationisemployedtoreducetheindeterminationdegree oftheimagePNswhichisevaluatedbytheentropyofthe indeterminate subset.Thus,theimagebecomesmoreuniform and homogenous and more suitable to be thresholded. Wethen usetheadaptivethresholdingmethodof[10]toobtainthe binaryimage. C.Post-prcessing Adaptive thresholding producesanoisybinaryimage.Consequently,apost-processingstageisrequiredtoremovethe noise.Forthis,amedianflterisusedwithawindowsizeof 3 x3"toenhancethebinaryimage. D.Summar Ourproposeddocumentanalysisalgorithmissummarized inAlgorithmI.ZAlgorithm1Proposeddocumentanalysisalgorithm I:ReadinRGBimagePRGB(X,y). 2:ConverttograyimagePg(x, y)usingNTSCstandard. 3:ApplyWieneradaptiveflteraspre-processingstepto obtainPgw(x, y). 4:TransformPgw(x,y)intoneutrosophicdomaintoobtain PNs(x, y) ={ T(x,y), I(x, y), F(x,y)} accordingtothe entropyofPgw(x,y)anditsmean. 5:MeasuretheentropiesofthethreesubsetsT,I,andF. 6:ApplyaA-meanoperationonPNS(X,y)todecreaseits indetermination. 7:SegmentthetruesubsetT usinganadaptivethresholding technique. 8:Applyamedianfltertoremovenoiseaspost-processing step. V.EXPERIMENTALRESULTS Inthissection,theproposedmethodisevaluatedand comparedwithotherbinarizationmethodsfromtheliterature. Ourdatasetcontainssamplescollectedfromboththe databaseof[2]andfromtheelectronicArabicmanuscripts of[16].Inourevaluation,wefocusonimagesthathave severaldegradations suchasmulti-coloredtextlines,stainsin thebackground, degraded charactersandmarks,and character diacritics. Forevaluationofourresults,groundtruthimagesare generatedusingasimilarmethodtotherecentworkof[17]. Wecomparetheperformanceofourproposedalgorithm withthoseobtainedbyotherbinarizationmethods,namely Otsu's[8],Niblack's[9],Sauvola's[10],andGuo's[5].The latterusesasimilarneutrosophicapproachbutthebinary imageisobtainedusingglobalthresholding. Fig.3demonstratestheresultsobtainedforanexample imageofthedataset. _. ...I " \|1M a) OriginalImage I'''',:-\'b ' ~e'~>-i,uG=' -'''~ .l!"I 1 .1 c) Savoula's , - b) Ground TruthImage _ _. ,1.....1 d)Y- Guo Fig.3.Exampleresultsobtainedbytheditferentbinarizationalgorithms. Ascanbeseen,ouralgorithmoutperformstheother approachesintermsofpreservationofmeaningfultextual information.Theothermethodseitherfailtosegmentthe foregroundtext,especiallyintheregionofstains(Fig.3(d), (e),and(f,orsegmentforegroundfrombackgroundbut addexcessivenoise(Fig.3(f.Ontheotherhand,ourfnal binaryimagesuffersfromsomestroke-likepatternnoise (SPN),whichisduetotheusedArabicmanuscriptsincluding diacritics.SPN[18],[19] issimilartodiacritics,andhenceits presenceneartextual components maychange themeaning of aword. Fortheimagesofourdataset,theaverageprocessingtime for our method was about 0.4 seconds for images of an average sizeof500 x 500 (on an Intel Corei3-2310- CPU@2.10GHZ with3GBRAM3.00,runningWindows7andthealgorithm implementedusingMatlabR2009a). Severalmethodshavebeenpresentedfortheevaluationof document image binarization techniques,and can be classifed intothreemaincategories[1].Evaluationcanbeperformed byvisualinspectionbyoneormorehumanevaluators.Here, symbolsthatarebrokenorblurred,andlossofobjectsas wellasnoiseinbackground and foreground are used asvisual evaluationcriteria.Clearly,thisapproachistimeconsuming andsubjective.Evaluationbasedonopticalcharacterrecognition(OCR)performance,appliesOCRontheresultimage andusestheobtainedcharacterorwordrecognitionaccuracy. ApplyingOCRasanevaluationcriterionisnotpossiblefor ourexperimentsduetothelackofaneffcientcommercialsoftwareforrecognizinghandwrittenArabicmanuscript writing[20].Finally,directevaluationofthebinarization canbeperformedbytakingintoaccountthepixel-to-pixel correspondencebetweenthegroundtruthandthebinarized image.Forthis,severalmeasurescanbeemployed[21]-[23], ofwhichweusethefollowinginourtests: F-measure[24],defnedas F= 2 x Precision x Recall Precision + Recall (22) where TP and Recall = TP +FN' TP Precision = TP+FP (23) (24) Here,T P denotestruepositives,TNtruenegatives,FP falsepositive,andFNfalsenegatives,respectively. F;-measure,defnedastheweightedharmonicmean betweenprecisionandrecall F= (1+()2XTP ; (1 + ()2 x T P + (2 X FN+ FP' (25) PSNR,a similaritymeasurebetweentwoimages,defned as[21],[25] with PSNR= 10l0g10

, (26) MSE= L [h(m,n) -h(m,n)]2 MxN (27) Z9wherehand arethetwoimages,and!andNthe imagedimensions. Negative metric rate(NRM),which isbased on pixelwise mismatchesbetweenthegroundtruthandthebinarized image[26],andcombinesthefalsenegativerateNFN andthefalsepositiverateN F Pas NVN + N"p NRM=NF+NTlNpl+NTN 2 (28) Abetterbinarizationqualityischaracterisedbyalower NRMvalue. Distancereciprocaldistortion(DRD)metric,defnedas s 2DRDk DRD=_K_=_l _ NUBN (29) whereDRDkisthedistortionofthek-thfippedpixel andiscalculatedusinga5 x 5normalizedweightmatrix W Nr [29].DRDkequalstheweightedsumofpixelsin the5 x5blockofthegroundtruthGTthatdiferfrom the centered k-th fipped pixel at(x,y) in the binarization resultB 22 DRDk=_ _IGTk(i,j) - Bk(i,j)1 xWNr(i,j). i=-2 j=-2 (30) NUBNisthenumberofthenon-uniform(notallblack orwhitepixels)8 x8blocksintheGTimage[27]. TableIsummarizestheresults(averagesoveralldataset images) of the various binarization algorithms. As is clear from theobtainedmeasures,ourproposedapproachprovidesthe bestresultswithrespecttoallperformancemetrics. TABLEI PERFORMANCE COMPARISON OF ALL BINARIZATION METHODS methodFFpPSNRNRMDRD Niblack84.4788.6313.514.050.072 Otsu90.3690.7215.882.650.037 Sauvola94.7695.3420.96l.200.033 Guo90.4490.8916.532.600.037 Proposed95.5995.8723.571.010.028 VI.CONCLUSIONS In this paper,we have proposed ahybrid thresholding techniquefordegradedimagesofhistoricalArabicmanuscripts. Ourmethodcombinesaneutrosophicsetapproachwithan adaptive thresholdinging method. Experiment results show that theproposedmethodprovidesgoodbinarizationperformance forcompleximagesthatcontainvariouschallengesincluding stains,inkseeping,andcharacterswrittenbydifferentink colors.Futureworkwillfocusonremovalofstroke-like patternnoisetofurtherimprovethebinarization results. ACKNOWLEDGMENTS Thanks to Dr.Y. Guo,St. ThomasUniv.,USA,for his kind helpandusefuldiscussions. Z70REFERENCES [I]K.Nitrogiannis, N.Gatos, andI.Pratikakis, "PerformanceEvaluation methodologyforhistoricaldocumentimagebinarization", IEEETrans ImageProcessing,Vol.22(2),2013. [2]K.M. Amin,M.A. Ahmad,and A. Ali,"A NovelBinaizationAlgorithm forHistoricalArabicManuscriptsusingWaveletDenoising",Int.J.of ComputingandInf.Sciences,Vol.13(1),2013. [3]PStathis,E.Kavallieratou, andN.Papamakos,"AnEvaluationTechniqueforBinarizationAlgorithms", JournalofUniversalComputer Science,Vol.14(18),pp.3011-3030, 2008. [4]J.Mohan,VKrishnaveni,YGuo,"ANewNeutrosophicApproachof WienerFilteringforMRIDenoising", MeasurementScienceReview, Vol.13(4),pp.177-168,2013. [5]YCheng, Heng-Da, andYanhuiGuo, "Anewneutrosophicapproach toimagethresholding", NewMathematicsandNaturalComputation, Vol.4(3),pp.291-308,2008. [6]M-L.FengandY-P.Tan,"Adaptivebinaizationmethodfordocument imageanalysis",IEEEInt.Conf.onMultimediaandExpo,Vol.1,pp. 339-342, 2004. [7]S.Faisal, K.Daniel, B.Thomas, "Effcientimplementationoflocal adaptivethresholdingtechniquesusingintegralimages",Electronic Imaging, pp.681510-681510, 2008. [8]N.Otsu,"AThresholdSelectionMethodfromGray-LevelHistograms ",IEEETransactionsonSystems,ManandCybernetics, Vol.9(1),pp. 62-66,1979. [9]W.Niblack, "AnIntroductiontoDigitalImageProcessing", Prentice Hall,EnglewoodCliffs,1986. [10]J.Sauvola, M.Pietikainen, "Adaptivedocumentimagebinaization", PatternRecognition, Vol.33(2), pp.225-236, 2000. [11]K.Khurram, "ComparisonofNiblackinspiredBinaizationmethods forancientdocuments", IST/SPIEElectronicImaging, pp.72470U-72470U,2009. [12]YGuo, andH.D.Chenga, "Newneutrosophicapproachtoimage segmentation",AdvancesinMultimedia,Vol.42(5),pp.587-595,2009. [13]M.Zhang,"NovelApproachestoImageSegmentationBasedonNeutrosophicLogic",DoctoralDissertation,UtahStateUniversity, 2010. [14]B.Gatos, 1.Pratikakis, andS.Perantonis, "AnAdaptiveBinaization TechniqueforLowQualityHistoricalDocuments",DAS2004,LNCS 3163,pp.102-113,2004. [15]A.Jain,"FundamentalsofDigitalImageProcessing",Pr.Hall,1989. [16]http://www.wqf.me.com [17]H.Nafchi, S.Ayatollahi, R.Farahi, andM.Cheriet, "Aneffcient groundtruthingtoolforbinarizationofhistoricalmanuscripts", 12th Int.Conf.onDocumentAnalysisandRecognition, pp.807-811,2013. [18]A. Faahmand,A. Sarrafzadeh,and J. Shanbehzadeh,"Document Image NoisesandRemovalMethods",IMECS,pp.436-440,2013. [19]M.AgrawalandD.Doermann,"Stroke-IikePatternNoiseRemovalin BinayDocumentImages",11thInt.Conf.onDocumentAnalysisand Recognition, pp.17-21,2011. [20]F.Smarandache, "AUnifyingFieldinLogicsNeutrosophicLogic. Neutrosophy, NeutrosophicSet, NeutrosophicProbability", 3rdEd., AmericanResearchPress,2003. [21]B.Gatos,K.Ntirogiannis, and1.Pratikakis, "ICDAR2009document imagebinaizationcontest(DIBCO2009)", Int.Conf.onDocument AnalysisandRecognition, pp.1375-1382, 2009. [22]H.Lu,A.C.Kot,Y.Q.Shi,"Distance-ReciprocalDistortionMeasure forBinayDocumentImages", IEEESignalProcessingLetters, Vol. 11(2),pp.228-231,2004. [23]B.Su, S.Lu, andC.Lim, "ARobustDocumentImageBinarization TechniqueforDegradedDocumentImages", IEEETrans.onImage Processing,Vol.22(4),2013. [24]N.Chinch or,"MUC-4evaluationmetrics",4thMessageUnderstanding Conference, pp.22-29,1992. [25]H. Lu,A. C. Kot,andY.Q. Shi,"Distance-reciprocaldistortionmeasure forbinarydocumentimages",IEEE SignalProcessingLetters,Vol.11, No. 2,pp.228-231,2004. [26]J. Aguilera,H.Wildenauer,M. Kampel,M. Borg,D. Thirde,and J. Ferryman,"Evaluationofmotionsegmentationqualityforaircraftactivity surveillance",2ndJointIEEEInt.WorkshoponVisualSurveillanceand Performance Evaluationof Tracking & Surveillance,pp. 293-300,2005. [27]1.Pratikakis, B.GatosandK.Ntirogiannis, "ICDAR2011Document ImageBinarizationContest", InternationalConferenceonDocument AnalysisandRecognition, pp.1506-1510,2011.