深度学习在手写汉字识别中的应用综述_金连文

428Vol.42,No.820168ACTAAUTOMATICASINICAAugust,2016112113,,,...,,,,.,,,,,,,,,,..,2016,42(8):X¡XDOI10.16383/j.aas.2016.c150725ApplicationsofDeepLearningforHandwrittenChineseCharacterRecognition:AReviewJINLian-Wen1ZHONGZhuo-yao1YANGZhao2YANGWei-Xin1XIEZe-Cheng1SUNJun3AbstractHandwrittenChinesecharacterrecognition(HCCR)isanimportantresearch¯ledofpatternrecognition,whichhasattractedextensivestudiesduringthepastfourdecades.Withtheemergenceofdeeplearning,newbreakthroughprogressesofHCCRhavebeenobtainedinrecentyears.Inthispaper,wereviewtheapplicationsofdeeplearningmodelsinthe¯eldofHCCR.First,theresearchbackgroundandcurrentstate-of-the-artHCCRtechnologiesareintroduced.Then,weprovideabriefoverviewofseveraltypicaldeeplearningmodels,andintroducesomewidelyusedopensourcetoolsfordeeplearning.TheapproachesofonlineHCCRando²ineHCCRbasedondeeplearningaresurveyed,withthesummariesoftherelatedmethods,technicaldetails,andperformanceanalysis.Finally,furtherresearchdirectionsarediscussed.KeywordsDeeplearning,handwrittenChinesecharacterrecognition,convolutionalneuralnetworks,recurrentneuralnetworks,long-shorttermmemory,stackedauto-encoderCitationJinLian-Wen,ZhongZhuo-Yao,YangZhao,YangWei-Xin,XieZe-Cheng,SunJun.ApplicationsofdeeplearningforhandwrittenChinesecharacterrecognition:areview.ActaAutomaticaSinica,2016,42(8):X¡X(Opticalcharacterrecognition,OCR),2015-11-042016-04-18ManuscriptreceivedNovember4,2015;acceptedApril18,2016(61472144),(2014A010103012,2015B010101004,2015B010130003,2015B010131004)SupportedbyNsationalNaturalScienceFoundationofChina(61472144),GDSTP(2014A010103012,2015B010101004,2015B010130003,2015B010131004)RecommendedbyAssociateEditorLIUCheng-Lin1.5106412.5106413.1001901.SchoolofElectronicandInformationEngineering,SouthChinaUniversityofTechnology,Guangzhou5106412.SchoolofMechanicalandElectricEngineering,GuangzhouUniver-sity,Guangzhou5106413.InformationTechnologyLabora-tory,FujitsuResearch&DevelopmentCenterCo.,Ltd,Beijing10019080,(HandwrittenChinesecharacterrecognition,HCCR),[1¡10].(O²ine)(Online).(),..,.,,,,2016-07-2214:49:07:1),1980GB2312-806763,.,2000GB18010(GB18010-2000),27533,;2005GB18010-200542711,70244.2),,,,,,,,;,,,;.;,,,;,,;,,.3),:\-\-\--\-\-\-\-\-\-,4),..HCL2000[11]863[12],;,CASIA-OLHWDB1.0-1.2[10],CASIA-HWDB1.0-1.2[10],SCUT-COUCH[13],GB2312-806763,.,:1)[14¡15][16][17¡19]([20¡22]);2),.,,[23¡25].HCCR,Gabor[26]Gradient[27];HCCR,8[24];3)(Modi¯edquadraticdiscriminantfunction,MQDF)[9;28](Supportvectormachine,SVM)[29](Hiddenmarkovmodel,HMM)[30](Discriminativelearningquadraticdiscriminantfunction,DLQDF)[31](Learningvectorquantity,LVQ)[32][9¡12],,,[33¡36],,[34],,Bayes[35],.,HCCR,[10](Discriminativefeaturelearning,DFE)(Discriminativelearn-ingquadraticdiscriminantfunction,DLQDF),CASIA-OLHWDBCASIA-HWDB[10],95.28%(DB1.0,4037)94.85%(DB1.1,3926)95.31%(ICDAR2013CompetitionDB,3755),94.20%(DB1.0),92.08%(DB1.1)92.72%(ICDAR2013Com-petitionDB).,,.[37]:,90%,95%,,//,,8:3.,.,\++,.,,,2011,ICDAR(InternationalConfer-enceonDocumentAnalysisandRecognition)[38¡39].,2013ICDAR[39],Graham(Spatially-sparseConvolutionalneuralnetwork)[40],,97.39%,CNN(Convolutionalneuralnetwork),,94.77%[39],HCCR,,,.,,,,,.[41¡43],(CNN)[44¡45](Deepbeliefnetwork,DBN)[41](Stackedauto-encoder,SAE)[46](Deeprecurrentneuralnetworks,DRNN)[47][48¡59],[40;60¡82],.,.:1;2;3.1[43;55],40[83],CNN,8090LeCun[44¡45;84],2006HintonScienceDBN[41],,DNNCNN[52][48;85],.,2012,[86¡89],,[52][44¡45;48¡59;85¡86;90][91¡92][93][60¡82;94¡95][96¡97][98¡102].,,,,,.,,.(DBN)S(SAE)(CNN)(Recurrentneuralnetwork,RNN),.(DBN)Hinton2006[41],,,,.,,[48;103][48][104],,.CNNFukushima1980[105],LeCun(Back-propagation,BP),[44¡45;84].(Convolutionallayer)(Poolinglayer),(Featuremap).2012,Krizhevsky[48]CNN,ReLUDropout,2012ImageNet[90].ReLUDropoutCNN,:1)ReLU:,f4XXx:f(x)=tanh(x)=ex¡e¡xex+e¡xf(x)=sigmoid(x)=11+e¡x.Krizhevsky[48]f(x)=max(0;x),ReLU.,sigmoidtanh,0,,,.ReLU0,,.2)Dropout:DropoutKrizhevsky[48],.Dropout(),0,,0,.,,.,Dropout,.,,,.DBN,Ranzato2007[46].SAEDBNRBM(Auto-encoder,AE),,,.RNN[94],DBNCNNSAE,,,().,RNN,(Backpropagationthroughtime,BPTT)[106].RNN,Hochre-iterSchmidhuber1997RNN(Longshorttermmemory,LSTM)[47].RNN,,.,(Deepre-inforcementnetwork,DRN)[107¡108],GoogleDeepMind2015Q(DeepQnetwork,DQN)[108],.,,DRN.,.Con-vNetA.KrizhevskyGPUCNN,2012ImageNet(LargeScaleVisualRecognitionChallenge2012,ILSVRC2012)[90],GPUCuda-ConvNet2[109];,,Ca®eTheanoTouchTensorFlow.1.2(),(CNN),(End-to-end),.,,CNN,,CNN.,,CNN..2.1CNNCNN90,()[44¡45;84;110¡111],,LeCun1998LeNet5CNN[45]MNIST99.05%,99.2%;2003,Simard[112](Elas-ticdistortion)(A±nedistortion)(Dataargumentation),CNN,MNIST,99.6%,SVMBoosting(MLP).CNN,,CPU.2011IDSIAGPU(GTX580)CNN,(1000)8:51Table1Somemainstreamdeep-learningopensourcetoolboxesandtheirdownloadaddressatpresentCa®e[112]UCBerkeleyBVLC,®eTheano[113¡114]Python[115]Lua,iOSAndroid[116]GPU,[117](DMCL)C++[118]NVID