Deep-Convolutional-Network-for-Handwritten-Chinese

DeepConvolutionalNetworkforHandwrittenChineseCharacterRec[e-mail]weexploredtheperformanceofdeepcon-volutionalneuralnetworkonrecognizinghandwrittenChi-nesecharacters.Weranexperimentsona200-classanda3755-classdatasetusingconvolutionalnetworkswithdif-ferentdepthandfilternumbers.Experimentalresultsshowthatdeepernetworkwithlargerfilternumbersgivebettertestaccuracy.WealsoprovideavisualizationofthelearnednetworkonthehandwrittenChinesecharacters.1.IntroductionDeepconvolutionalneuralnetwork(CNN)hasbecomethearchitectureofchoiceforcomplexvisionrecognitionproblemsforseveralyears.TherehasbeenalotofresearchonusingdeepCNNtorecognizehandwrittendigits,En-glishalphabets,orthemoregeneralLatinalphabets.Ex-perimentshaveshownthatwell-constructeddeepCNNsarepowerfultoolstotacklethesechallenges.Astherecogni-tionofcharactersinvariouslanguageshasattractedmuchattentionintheresearchcommunity,anaturalquestionis:HowdoesdeepCNNperformforrecognizingmorecom-plexhandwrittencharacters?Inthisproject,wewillexplorethepowerofdeepCNNontheclassificationofhandwrittenChinesecharacters.ComparedtothetaskofrecognizinghandwrittendigitsandEnglishalphabets,therecognitionofhandwrittenChi-nesecharactersisamorechallengingtaskduetovariousreasons.Firstly,therearemuchmorecategoriesforChi-nesecharactersthanfordigitsandEnglishcharacters.Asacomparison,thereare10digitsforusualdigitrecognitiontasks,andthereare26alphabetsforEnglish,whilethereareintotalover50,000Chinesecharactersandaround3,000ofthemareforeverydayuse.Secondly,mostChinesechar-actershavemuchmorecomplicatedstructuresandconsistofmuchmorestrokescomparedtodigitsorEnglishcharac-ters.Figure1showsacomparisonofdifferenthandwrittencharacters.Thirdly,handwritingstyleforChinesecharac-tersvarieshugelyfrompersontoperson.Moreover,theexistenceofjoined-uphandwritingmakestherecognitionevenmoredifficult.Forexample,Figure2showsthein-fluenceofdifferenthandwritingstylesontheappearanceofhandwrittenChinesecharacters.Itisevenachallengingtaskforawell-educatedChinesetorecognizeallthehand-writtencharacterscorrectly.Inthisproject,wewillfocusontwospecificquestions:1)Howwillthearchitectureanddepthinfluencetheaccu-racyofCNNonrecognizinghandwrittenChinesecharac-ters?2)Doestheextractedfeaturesmakesenseintermsofvisualization?Therestofthereportisorganizedasfollows.Wewillfirstintroducedthedatasetandournetworkconfig-urationsinSection2andSection3.ThenwewillintroducehowweimplementandtrainournetworksinSection4.Af-terwardswepresentourexperimentalresultsinSection5andanalyzetheresultsinSection6.Finally,wewilldis-cusstherelatedworkinSection7.(a)Digit(b)English(c)ChineseFigure1:Exampleofhandwrittencharacters2.Data2.1.DatasetForthisprojectweusetheCASIAofflinedatabase,asdescribedin[6].Thedataconsistsofplaingray-scaleim-agesofisolatedhandwrittenChinesecharacters,asshowninFigure2.Specifically,wewillusetheHWDB1.1dataset,whichtotallyincludes3,755Chinesecharactersand171al-phanumericandsymbols.AsispresentedinTable1,eachcategorycontainshandwrittenimagesfromapproximately300writers(withminordifferenceforsomecategories),andeachwritercontributesoneimagetoeachcategory.Asre-leased,thefulldatasetissplitintotwoparts:atrainingset1Figure2:Differentdataexamplesinthesamecategory.ExamplesoneachrowcomefromdifferentwritersandcorrespondtothesameChinesecharacter:艾,斌,and棉respectively.Verydifferenthandwritingstylesacrosswriterscouldbeobserved.Dataset#Writers#Classes#TotalSamples#Chinesecharacters#SymbolsCASIAHWDB1.13003,7551,172,9071,121,74951,158Table1:HWDB1.1DatasetInformationandtestset.Testsetcontains60randomlysampledimagesforeachcategory,andtrainingsetcontainstherest(approx-imately240).Inthisproject,fordebuggingandcomparingdifferentmodelsduringtraining,wefurthersplittheorigi-naltrainingsetintotwoparts:atrainingsetandavalidationset,withtrainingsetcontaining200imagesforeachcate-goryandvalidationsetcontainstherest(approximately40).Trainingonthefulldataset(over1millionexamples)cantakemanyhoursordaysevenwiththefastestGPU.Constraintbythecomputationresourceswehaveaccessto,weranourmajorexperiments(formodelcomparisonandvisualization)onasubsetofthefulldataset,whichcontains200randomlysampledclasses.Thesizeoftrainingset,val-idationsetandtestsetforeachclassremainsunchanged.Wealsoranexperimentsonthefulldatasettoevaluatethepowerofourbestmodels.Thisfulldatasetcontainsallthe3,755classes.Itisworthnotingthattherearemuchmoreclassesthanexamplesforeachclassesinthetrainingset.TheinformationabouttwodatasetsarelistedinTable2.2.2.PreprocessingThereleaseddatasetcontainsexamplesinbinaryformat,alongwithlabels.Sothefirststepofthedataprocessingistoconvertthebinarydataintoimageformat.Hereweuse.jpgtoencodetheimageandstoretheimagefiles.Theconvertedimageshavebackgroundlabeledas255andfore-groundpixelsin255graylevels(0-254).Asitisusedin[2],herewefollowedathree-stepprepro-cessingapproach:resizing,contrastmaximizationandim-agemeansubtraction.GivenarawinputimagedescribingahandwrittenChinesecharacter,wefirstresizedtheimageintoanormalizeds