基于卷积神经网络的手写孟加拉语和孟加拉语混合数字识别(IJIGSP-V8-N9-6)

I.J.Image,GraphicsandSignalProcessing,2016,9,40-50PublishedOnlineSeptember2016inMECS()DOI:10.5815/ijigsp.2016.09.06Copyright©2016MECSI.J.Image,GraphicsandSignalProcessing,2016,9,40-50ConvolutionalNeuralNetworkbasedHandwrittenBengaliandBengali-EnglishMixedNumeralRecognitionM.A.H.Akhand,MahtabAhmedDept.ofComputerScienceandEngineering,KhulnaUniversityofEngineering&Technology(KUET)Khulna-9203,BangladeshEmail:{akhand,mahtab}@cse.kuet.ac.bdM.M.HafizurRahmanDept.ofComputerScience,KICT,InternationalIslamicUniversityMalaysia(IIUM)Selangor,MalaysiaEmail:[e-mail]—Recognitionofhandwrittennumeralshasgainedmuchinterestinrecentyearsduetoitsvariouspotentialapplications.Bengaliisthefifthrankedamongthespokenlanguagesoftheworld.However,duetoinherentdifficultiesofBengalinumeralrecognition,averyfewstudyonhandwrittenBengalinumeralrecognitionisfoundwithrespecttoothermajorlanguages.TheexistingBengalinumeralrecognitionmethodsuseddistinctfeatureextractiontechniquesandvariousclassificationtools.Recently,convolutionalneuralnetwork(CNN)isfoundefficientforimageclassificationwithitsdistinctfeatures.Inthispaper,wehaveinvestigatedaCNNbasedBengalihandwrittennumeralrecognitionscheme.SinceEnglishnumeralsarefrequentlyusedwithBengalinumerals,handwrittenBengali-Englishmixednumeralsarealsoinvestigatedinthisstudy.Theproposedschemeusesmoderatepre-processingtechniquetogeneratepatternsfromimagesofhandwrittennumeralsandthenemploysCNNtoclassifyindividualnumerals.Itdoesnotemployanyfeatureextractionmethodlikeotherrelatedworks.TheproposedmethodshowedsatisfactoryrecognitionaccuracyonthebenchmarkdatasetandoutperformedotherprominentexistingmethodsforbothBengaliandBengali-Englishmixedcases.IndexTerms—ImagePre-processing,ConvolutionalNeuralNetwork,BengaliNumeral,HandwrittenNumeralRecognition.I.INTRODUCTIONRecognitionofhandwrittennumeralshasgainedmuchinterestinrecentyearsduetoitsvariouspotentialapplications.Theseincludepostalsystemautomation,passports&traveldocumentanalysis,automaticbankchequeprocessing,andevenfornumberplateidentification[1,2].ResearchonrecognitionofunconstrainedhandwrittennumeralshasmadeimpressiveprogressinRoman,Chinese,andArabicscript[3,4].Ontheotherhand,recognitionofhandwrittenBengalinumeralislargelyneglectedalthoughitisamajorlanguageinIndiansubcontinenthavingfifthrankedintheworldandisthemainlanguageofBangladesh.SomeBengalinumeralshaveverysimilarshape,andevenintheprintedformfewdiffersverylittlethanothernumerals.Therefore,Bengalinumeralrecognitionisachallengingtask.AninterestingaspectofBengalidocumentsisthatEnglishentriesarecommonlyavailableinthose.TheseincludethetextbookswritteninBengaliscriptsoftenhaveentriesinEnglishespeciallythenumerals;BangladeshicurrenciescontainbothBengaliandEnglishnumeralstorepresentvalues;handwrittenBengali-EnglishmixednumeralsarefrequentlyfoundinBangladeshwhilewritingpostalcode,bankcheque,age,numberplate,mobilenumberandothertabularformdocuments.Moreover,oftenpeoplecasuallyenteroneormoreEnglishnumeralswhichresultsamixed-scriptsituation.ThereisasimilarityinwritingstyleofseveralBengalinumerals(e.g.,‗০‘,‗২‘,‘৪‘and‗৭‘)withEnglishnumerals(e.g.,‗0‘,‘2‘,‗8‘and‗9‘)thatmakesBengali-Englishmixednumeralrecognitionmorechallenging.ThereasonbehindtheusesofthismixedcaseisthatEnglishisusedinparallelwithBengaliinofficialworksandeducationsystems.Therefore,Bengali-Englishmixednumeralrecognitionsystemischallengingandimportantforpracticalapplications.AlthoughseveralremarkableworksareavailableforBengaliandEnglishhandwrittennumeralrecognitionseparately,averyfewworksareavailableformixedBengali-Englishnumeralrecognitionwhoseperformancearenotatsatisfactorylevel.Therestofthepaperisorganizedasfollows.SectionIIreviewsseveralrelatedworksandexplainsmotivationofthepresentstudy.SectionIIIexplainsproposedrecognitionschemeusingconvolutionalneuralnetwork(CNN)whichcontainsdatasetpreparation,pre-processingandclassification.SectionIVpresentsConvolutionalNeuralNetworkbasedHandwrittenBengaliandBengali-EnglishMixedNumeralRecognition41Copyright©2016MECSI.J.Image,GraphicsandSignalProcessing,2016,9,40-50experimentalresultsoftheproposedmethodandcomparisonofperformancewithotherrelatedworks.Finally,abriefconclusionoftheworkisgiveninSectionV.II.RELATEDWORKSAfewnotableworksareavailableforBengalihandwrittennumeralrecognitionwithrespecttootherpopularIndiansubcontinentscriptssuchasDevanagari[5-7].Basharetal.[8]investigatedadigitrecognitionsystembasedonwindowingandhistogramtechniques.Windowingtechniqueisusedtoextractuniformfeaturesfromscannedimagefilesandthenhistogramisproducedfromthegeneratedfeatures.Finally,recognitionofthedigitisperformedonthebasisofgeneratedhistogram.Paletal.[3]introducedanewtechniquebasedontheconceptofwateroverflowfromthereservoirforfeatureextractionandthenemployedbinarytreeclassifierforhandwrittenBengalinumeralrecognition.Basuetal.[4]usedDempster-Shafer(DS)techniquetocombinetheclassificationdecisionsobtainedfromtwoMLPbasedclassifiersforhandwrittenBengalinumeralrecognitionusingtwodifferentfeaturesets.Fea