关于车牌识别的英文文献

CharacterRecognitioninCarLicensePlatesBasedonPrincipalComponentsandNeuralProcessingAlinedaRochaGesualdi,Jos´eManoeldeSeixasRiodeJaneiroFederalUniversitySignalProcessingLaboratory(LPS),EletricEngineeringDepartament,COPPE/UFRJRiodeJaneiro-RJ,Brazil,CP68504,[e-mail];[e-mail]rque,M´arcioPortesdeAlbuquerqueBrazilianCenterforResearchinPhysics(CBPF)TechnicalActivityDepartment(CAT)RiodeJaneiro,RJ,[e-mail],[e-mail],indetails,thePrincipalComponentAnalysis(PCA)appliedonaneuralbasedrecognitionsystem.Thetechniqueisvaluatedontheextractionofprincipalcompo-nentsincharacterimagesofBrazilians’licenseplates.ThispaperfocusontheusageofthePCAintherecognitionofcharacterimages.Comparisonswithdifferentneuralclas-sifiersaremade.1.IntroductionThedesignofintelligenttransportationsystemsforap-plicationssuchaselectronictoll,trafficflowanalysis,speedlimitandredlightviolationenforcementhasbeenattractedanincreasingattentionfromresearchers.Inparticular,therecognitionofcharactersincarlicenseplatesisobjectofdifferentstudiesnowadays.Asaverypowerfultechniqueforpatternrecognitionproblems,itisnotsurprisingthatneuralnetworksbecamethebasisofanumberofsuchsys-temdesigns[8,2,1].Recently,asuccessfuldesignforBraziliancarplateshasbeenreported[3].Thedesignapproach,whichissum-marizedhereforreference,includedanumberofprepro-cessingstepsaimingatextractingefficientlythecharactersegmentsfromdigitalimageplates.Splittingtheclassifi-cationtaskintotwo,twoexpertneuralclassifiers[5]weredesigned,onefocusingontheidentificationofdigitsandanotheronletters.Bothneuralnetworksweretrainedwithbackpropagationmethod[4]andhadasinglehiddenlayer.Thenetworkswerefedwithspecificpreprocesseddata(oneformedwithnumbersandtheotherwithlettersextractedfromthecarplatesthroughmorphologicalimageprocess-ingtechniques)fromadatabasecomprising432platesandachievedaclassificationefficiencyforcharactersbetterthan99%.Theoutputlayerforeachexpertnetworkhadneu-ronsassignedtoclasses,sothatatotalof10outputneuronswerepresentfornumbersand26outputneuronswereusedforletters.Thecarplateidentificationwasestimatedtobeabove85%,whenmaximumoutputprobabilitywasconsid-ered.TheBrazilianslicenseplateshavethreelettersandfournumbers.In[3],360plateimageswereusedinordertotrainthetwoneuralnetworks,oneforthelettersandtheotherforthenumbers.Theremaining72plateimageswereusedtotesttherecognitionsystem.Theneuralnetworkspresentedin[3]had,asinputs,thecharacterimages,with256graylevels.Theseimageswereobtainedbythemanualextractionofthecharactersintheplateimages,theneachcharacterimagewasresizedtopixelimagesandre-shapedtovectors.Afterthisprocess,thepatternsweresep-aratedingroupsoflettersandnumberstobepresentedtotheneuralnetworksasinputpatterns.Thestructuresoftheneuralnetworksweredesignedusingaconstructivealgorithm;wherethebasicideawastostartwithasmallnetwork,thenaddhiddenunitsandweightsincrementallyuntilasatisfactorysolutionbefound.Thehigherefficiencieswereforanumbernet-workandletternetwork.Inthispaper,aprincipalcomponentanalysisisper-formedontherandomprocessdefinedbyplateimages,ProceedingsoftheVIIBrazilianSymposiumonNeuralNetworks(SBRN’02)0-7695-1709-9/02$17.00©2002IEEEwhichwillbethesameimagesusedinthedatabasereferredtoin[3].Themainmotivationforusingprincipalcom-ponentsistoevaluatewhethertherelevantfeaturesfortheclassificationtaskcanstillbeextractedfromareducednum-berofcomponents,which,inaddition,mayreducethecom-plexityoftheneuralclassifiers.Forthis,twoapproachesarefollowed.Firstly,theclassicallow-rankprincipalcom-ponentanalysis(PCA)[7]isperformed,envisagingtopre-servemostoftheenergyoftheoriginalinformationandstillachievingagoodcompressionratefortheoriginalin-putdataspace.Thesecondapproachmakesuseoftheprin-cipalcomponentdiscrimination(PCD)method,whichdoesnottrytoreconstructtheoriginalinputdataspacewithareducednumberofcomponentsbuttriestoidentifythebestcomponentsforperformingtheclassificationtask[6].Typ-ically,PCDachievesahighercompressionrateincompari-sontoPCA.PCAandPCDwerebothperformedonthetrainingsetusedfortheneuralnetworkdesigns.Componentswereex-tractedfromtheset,previouslymentioned,thatcomprises360images.Thetrainingsetwassplitintotwotoformthevalidationset,whichwasusedforvalidatingthetrainingphaseandavoidingovertrainingeffects[4].Thesetformedbythose72imageswereleftoutofthetrainingphaseandwasusedtoevaluatethegeneralizationcapacityofthepro-posedsystem,formingthesocalledtestingset.Asin[3]pixelintensitiesfromcharactersweremeasuredusing256graylevelsandcharacterimageswereresizedtopixelsandreshapedtovectors.Thenextsectionconcernstheprincipalcomponentanal-ysisonimageplatesandtheclassificationefficiencyob-tainedwhendatacompressionthroughPCAisrealized.SectionIIIdetailstheprincipalcomponentdisciminationmethodandevaluatesitsabilityforthisapplication.Finally,SectionIVcomparesclassificationresultsandderivessomeconclusions.2.EvaluationofPCAdefinitionThelow-rankprincipalcomponentanalysisisbasedonevaluatingtheenergystoredoneachcomponentoftheKarhunen-Lo