基于支持向量机和人工神经网络的图像分类(IJITCS-V4-N5-5)

I.J.InformationTechnologyandComputerScience,2012,5,32-38PublishedOnlineMay2012inMECS()DOI:10.5815/ijitcs.2012.05.05Copyright©2012MECSI.J.InformationTechnologyandComputerScience,2012,5,32-38ImageClassificationusingSupportVectorMachineandArtificialNeuralNetworkLeHoangThaiComputerScienceDepartment,UniversityofScience,HoChiMinhCity,VietnamEmail:[e-mail]echnologyDepartment,UniversityofPedagogy,HoChiMinhCity,Vietnam,memberofIACSITEmail:[e-mail]ityofTechnology,HaNoiCity,VietnamEmail:[e-mail]—Imageclassificationisoneofclassicalproblemsofconcerninimageprocessing.Therearevariousapproachesforsolvingthisproblem.TheaimofthispaperisbringtogethertwoareasinwhichareArtificialNeuralNetwork(ANN)andSupportVectorMachine(SVM)applyingforimageclassification.Firstly,weseparatetheimageintomanysub-imagesbasedonthefeaturesofimages.Eachsub-imageisclassifiedintotheresponsiveclassbyanANN.Finally,SVMhasbeencompiledalltheclassifyresultofANN.OurproposalclassificationmodelhasbroughttogethermanyANNandoneSVM.LetitdenoteANN_SVM.ANN_SVMhasbeenappliedforRomannumeralsrecognitionapplicationandtheprecisionrateis86%.Theexperimentalresultsshowthefeasibilityofourproposalmodel.IndexTerms—imageclassification,supportvectormachine,artificialneuralnetwork1.IntroductionImageclassificationisoneofclassicalproblemsofconcerninimageprocessing.Thegoalofimageclassificationistopredictthecategoriesoftheinputimageusingitsfeatures.Therearevariousapproachesforsolvingthisproblemsuchasknearestneighbor(K-NN),Adaptiveboost(Adaboosted),ArtificialNeuralNetwork(NN),SupportVectorMachine(SVM).Thek-NNclassifier,aconventionalnon-parametric,calculatesthedistancebetweenthefeaturevectoroftheinputimage(unknownclassimage)andthefeaturevectoroftrainingimagedataset.Then,itassignstheinputimagetotheclassamongitsk-NN,wherekisaninteger[1].Adaboostedisafastclassifierbasedonthesetofweakclassifiers.AweakclassifierbasedonHaar-Likefeaturescouldbedefined[2]as:1()0jjjjifpfxphotherwise(1)Wherexisasub-window,andθisathreshold.pjindicatingthedirectionoftheinequalitysign.AdaBoost(AdaptiveBoost)isaniterativelearningalgorithmtocreatea“strong”classifierusingatrainingdatasetanda“weak”learningalgorithm.Ateveryiterativestep,the“weak”classifierwiththeminimumclassificationerrorisselected.ArtificialNeuralNetwork(ANN),abrain-stylecomputationalmodel,hasbeenusedformanyapplications.ResearchershavedevelopedvariousANN’sstructureinaccordantwiththeirproblem.Afterthenetworkistrained,itcanbeusedforimageclassification.SVMisoneofthebestknownmethodsinpatternclassificationandimageclassification.Itisdesignedtoseparateofasetoftrainingimagestwodifferentclasses,(x1,y1),(x2,y2),...,(xn,yn)wherexiinRd,d-dimensionalfeaturespace,andyiin{-1,+1},theclasslabel,withi=1..n[1].SVMbuildstheoptimalseparatinghyperplanesbasedonakernelfunction(K).Allimages,ofwhichfeaturevectorliesononesideofthehyperplane,arebelongtoclass-1andtheothersarebelongtoclass+1.BesidestherearesomeintegratedmultitechniquesmodelforclassifyingsuchasMultiArtificialNeuralNetwork(MANN)applyingforfacialexpressionclassification,andMultiClassifierSchemeapplyingforAdultimageclassification.MANNmodelareshowninthefollowingdiagram:ImageClassificationusingSupportVectorMachineandArtificialNeuralNetwork33Copyright©2012MECSI.J.InformationTechnologyandComputerScience,2012,5,32-38Fig.1MultiArtificialNeuralNetworkmodel[3]IntheaboveFig.1,MultiArtificialNeuralNetwork(MANN)[4],applyingforpatternorimageclassificationwithparameters(m,L),hasmSub-NeuralNetwork(SNN)andaglobalframe(GF)consistingLComponentNeuralNetwork(CNN).Inparticular,misthenumberoffeaturevectorsofimageandListhenumberofclasses.ThismodelusesmanyNeuralNetworkssothatthetrainingphraseiscomplexandlong.Besides,itisnotsuitableincasethenumberofclassesLishigh.MANNisthe2-layersclassifiermodelusingNeuralNetwork.BesidesmulticlassifierschemehasjustbeenproposedforAdultimageclassificationwithlowlevelfeaturein2011[5].Thismodelcontainstwo-layersclassifier.Layer1usesSupportVectorMachine(SVM)classifierandAdaBoostclassifier.Layer2isthemajoritybaseclassifierintegratingtheclassifiedresultsoflayer1.MultiClassifierSchememodelisshowninthefollowingdiagram:Fig.2MultiClassifierSchememodel[5]IntheaboveFig.2,theMultiClassifierSchememodelistwolayersclassifier.TheoutputofSVMclassifierandAdaBoostclassifierhasbeencombinedbyMajorityBaseClassifier.Thisexperimenthasshowedthatweneedtochoosetheappropriateclassifiersforthefeatureextractiontoincreasetheprecisionofimageclassification.Ontheotherhand,theprecisionofclassificationsystemdependsonthefeatureextractionandtheclassifier.Theremainderofthispaperisorganizedasfollows:Section2devotedtostudyofimageclassificationprocessanditsproblems.Section3providesadetailedexpositionofourproposalmodelANN_SVMwhichhasbeencompiledmanyArtificialNeuralNetworksandtheSupportVectorMachine.Section4containsadiscussionoftheexperimentsandevaluationofRomannumeralrecognitionapplicationusingourproposalmodelANN_SVM.Conclusionandf