基于卷积神经网络的BangLA手写体字符识别(IJIGSP-V7-N8-5)

I.J.Image,GraphicsandSignalProcessing,2015,8,42-49PublishedOnlineJuly2015inMECS()DOI:10.5815/ijigsp.2015.08.05Copyright©2015MECSI.J.Image,GraphicsandSignalProcessing,2015,8,42-49BanglaHandwrittenCharacterRecognitionusingConvolutionalNeuralNetworkMd.MahbubarRahman,M.A.H.Akhand,ShahidulIslam,PintuChandraShillDept.ofComputerScienceandEngineeringKhulnaUniversityofEngineering&TechnologyKhulna,BangladeshEmail:{akhand,pintu}@cse.kuet.ac.bdM.M.HafizurRahmanDept.ofComputerScience,KICT,InternationalIslamicUniversityMalaysiaSelangor,MalaysiaEmail:[e-mail]—Handwrittencharacterrecognitioncomplexityvariesamongdifferentlanguagesduetodistinctshapes,strokesandnumberofcharacters.NumerousworksinhandwrittencharacterrecognitionareavailableforEnglishwithrespecttoothermajorlanguagessuchasBangla.Existingmethodsusedistinctfeatureextractiontechniquesandvariousclassificationtoolsintheirrecognitionschemes.Recently,ConvolutionalNeuralNetwork(CNN)isfoundefficientforEnglishhandwrittencharacterrecognition.Inthispaper,aCNNbasedBanglahandwrittencharacterrecognitionisinvestigated.TheproposedmethodnormalizesthewrittencharacterimagesandthenemployCNNtoclassifyindividualcharacters.Itdoesnotemployanyfeatureextractionmethodlikeotherrelatedworks.20000handwrittencharacterswithdifferentshapesandvariationsareusedinthisstudy.Theproposedmethodisshownsatisfactoryrecognitionaccuracyandoutperformedsomeotherprominentexitingmethods.IndexTerms—HandwrittenCharacterRecognition,Bangla,ConvolutionalNeuralNetwork.I.INTRODUCTIONInrecentyears,therehasbeenmuchinterestinautomaticcharacterrecognition.Betweenhandwrittenandprintedforms,handwrittencharacterrecognitionismorechallenging.Handwrittencharacterswrittenbydifferentpersonsisnotidenticalbutvariesinbothsizeandshape.Numerousvariationsinwritingstylesofindividualcharactermaketherecognitiontaskdifficult.Thesimilaritiesindistinctcharactershapes,theoverlaps,andtheinterconnectionsoftheneighboringcharactersfurthercomplicatetheproblem.Ahandwrittencharacterrecognitionsystemconsistsoftwomajorsteps:featureextractionfromthecharactersetandthenemploylearningtool(s)toclassifyindividualcharacter[1-4].Withdistinctfeatureextractiontechniques,anumberofmethodsbasedonartificialneuralnetworkareinvestigatedforhandwrittenEnglishcharacterrecognition.Morphological/Rank/LinearNeuralNetwork(MRL-NN)[1],inwhichthecombinationofinputsineverynodeisformedbyhybridlinearandnonlinear(ofthemorphological/ranktype)operations,isinvestigatedforhandwrittendigit.AhybridMultilayerPerceptron-SupportVectorMachine(MLP-SVM)basedmethodwasusedforEnglishdigit[2]andChinesecharacter[3]recognition.SupportVectorMachine(SVM)withRadialBasisFunction(RBF)networkisusedinRef.[4]forbothsmallandcapitalhandwrittenEnglishcharacterset.Recently,ConvolutionalNeuralNetwork(CNN)[5]isfoundefficientforhandwrittencharacterrecognitionduetoitsdistinctfeatures.CNNsaddthenewdimensionforimageclassificationsystemsandrecognizingvisualpatternsdirectlyfrompixelimageswithminimalpreprocessing.Inaddition,CNNautomaticallyprovidessomedegreeoftranslationinvariance.ACNNbasedmodelwastestedonUNIPEN[6]Englishcharacterdatasetandfoundrecognitionratesof93.7%and90.2%forlowercaseanduppercasecharacters,respectively[7].Moreover,CNNcommitteesarealsofoundtoenhancetherecognitionperformancewherethecommitteesareformedbyvariesaspectratiosoncharactersimagestosamearchitecture[8].Characterrecognitioncomplexityvariesamongdifferentlanguagesduetodistinctshapes,strokesandnumberofcharacters.Anumberofworks,includingtheabovediscussedmethods,isavailableforEnglishwithrespecttoothermajorlanguagessuchasBangla.Banglaisoneofthemostspokenlanguages,rankedfifthintheworld.Itisalsoanimportantlanguagewitharichheritage;21stFebruaryisdeclaredastheInternationalMotherLanguagedaybyUNESCOtorespectthelanguagemartyrsforthelanguageinBangladeshattheyearof1952.BanglaisthefirstlanguageofBangladeshandthesecondmostpopularlanguageinIndia.About220millionpeopleuseBanglaastheirspeakingandwritingpurpose.Thereare50charactersinBanglaandsomecontainsadditionalsignupand/orbelow.Moreover,Banglacontainsmanysimilarshapedcharacters;insomecasesacharacterdifferfromitssimilaronewithasingledotormark.ThatmakesdifficulttoachievebetterperformancewithsimpletechniqueaswellashinderstoworkwithBanglahandwrittencharacterrecognition.AfewnotableworksareavailableforBanglaBanglaHandwrittenCharacterRecognitionusingConvolutionalNeuralNetwork43Copyright©2015MECSI.J.Image,GraphicsandSignalProcessing,2015,8,42-49handwrittencharacterrecognition.Bhowmiketal.[9]proposedafusionclassifierusingMultilayerPerceptron(MLP),RBFnetworkandSVM.Theyusedwavelettransformforfeatureextractionfromcharacterimages.Inclassification,theyconsideredsomesimilarcharactersasasinglepatternandtrainedtheclassifierfor45classes.Basuetal.[10]proposedahierarchicalapproachtosegmentcharactersfromwordsandMLPisusedforclassification.Insegmentationstagetheyusedthreedifferentfeatureextractiontechniquesbuttheyreducedcharacterpatternsinto36classesmergingsimilarcharactersinasingleclass.Recent