基于卷积神经网络的图像分类优化体系结构(IJIGSP-V11-N10-5)

I.J.Image,GraphicsandSignalProcessing,2019,10,30-39PublishedOnlineOctober2019inMECS()DOI:10.5815/ijigsp.2019.10.05Copyright©2019MECSI.J.Image,GraphicsandSignalProcessing,2019,10,30-39AnOptimizedArchitectureofImageClassificationUsingConvolutionalNeuralNetworkMuhammadAamir1,ZiaurRahman11CollegeofComputerScienceSichuanUniversity,No.24SouthSection1,YihuanRoad,Chengdu,China,610065Email:[e-mail],[e-mail]22SchoolofComputerScienceandEngineeringSoutheastUniversitySipailouNo.2,Nanjing,China,210096Email:[e-mail]3,SyedMustajarAhmed43SchoolofSoftwareTechnology,4SchoolofComputerScienceandElectricalEngineeringDalianUniversityofTechnology,Dalian,China,116620Email:[e-mail],[e-mail]:22July2019;Accepted:23August2019;Published:08October2019Abstract—Theconvolutionalneuralnetwork(CNN)isthetypeofdeepneuralnetworkswhichhasbeenwidelyusedinvisualrecognition.Overtheyears,CNNhasgainedlotsofattentionduetoitshighcapabilitytoappropriatelyclassifyingtheimagesandfeaturelearning.However,therearemanyfactorssuchasthenumberoflayersandtheirdepth,numberoffeaturesmap,kernelsize,batchsize,etc.Theymustbeanalyzedtodeterminehowtheyinfluencetheperformanceofnetwork.Inthispaper,theperformanceevaluationofCNNisconductedbydesigningasimplearchitectureforimageclassification.WeevaluatedtheperformanceofourproposednetworkonthemostfamousimagerepositorynameCIFAR-10usedforthedetectionandclassificationtask.Theexperimentresultsshowthattheproposednetworkyieldsthebestclassificationaccuracyascomparedtoexistingtechniques.Besides,thispaperwillhelptheresearcherstobetterunderstandtheCNNmodelsforavarietyofimageclassificationtask.Moreover,thispaperprovidesabriefintroductiontoCNN,theirapplicationsinimageprocessing,anddiscussrecentadvancesinregion-basedCNNforthepastfewyears.IndexTerms—Convolutionalneuralnetwork,deeplearning,imageclassification,precision,recall.I.INTRODUCTIONCNNisoneofthemostpopulardeepneuralnetworkarchitectures,whichcomesinnumerousvariations.Thereareseveralfeaturessuchasconvolutionaloperation,characteristicofparametersharing,andshift-invariant,whichmakesthemtypicaldeeplearningmodelincomputervision.TheunderlyingCNNarchitecturehasformedbystackingthreetypesoflayersontopofeachother,thatareaconvolutionallayer[1],poolinglayer,andfullyconnectedlayer,alsoknownasadenselayer[2][3].AsimplifiedCNNarchitecturefordogclassificationisillustratedinFig.1.Moreover,CNNoffersexceptionalperformanceinmachinelearningproblems.Inparticular,applicationsthatdealwithimagedata,suchasacompleteimageclassificationdataset.Inthelastdecade,theaccuracyofimageclassificationhasbeenimprovedwiththeadvanceofdeeplearning,especiallyconcerningCNN.Thisincreaseintheefficiencyofimageclassificationhasledresearchersanddeveloperstoapproachlargermodelstosolvecomplexproblems,whichwasnotpossiblewithclassicalartificialneuralnetworks(ANNs)[4].CNNhasbeeneffectivelyappliedtoavarietyofdeeplearningproblems,suchasobjectrecognition,objectclassification,speechrecognition,and,inparticular,problemsassociatedwithmassiveimagedata.ThefirstCNNisintroducedbyLeCunetal.[5]in1990,anditsimprovedversiondevelopedin1998[6].Amulti-layerANNwhichcanbetrainedwiththebackpropagationalgorithmcalledLeNet-5hasbeendesignedtoclassifyhandwrittendigits[7].Thenetworkcantransformtheoriginalimageintousefulrepresentationssothatitcanrecognizevisualpatternsdirectlyfromunprocessedpixelswithoutmuchpreprocessing[8].However,duetothelackofextensivetrainingdataandcomputingpoweratthistime,LeNet-5cannotachieveexcellentresultsinmorecomplexmatters,suchas:Intheclassificationofimagesandvideosonalargescale.AfterthisnumberofmethodsAnOptimizedArchitectureofImageClassificationUsingConvolutionalNeuralNetwork31Copyright©2019MECSI.J.Image,GraphicsandSignalProcessing,2019,10,30-39Fig.1.AsimplifiedCNNarchitecturefordogClassificationhavebeenproposedtoaddressthedifficultiesintrainingthedeepneuralnetwork.MostunusualisapopulardeepCNN,developedbyKrizhevshyetal.[9]wasAlexNet.Thenetworkintroducedin2012includedtheavailabilityofhighcomputingdevices(i.e.,GPU,averydeepnetworkof60millionand650,000neurons,etc.).AlexNetoutperformedallpreviouscompetitorsandacceptedthechallengebyreducingtheerrorofthetop5to15.3%.Theerrorrateofthetop5positions,whichisnotavariationofCNN,wasaround26.2%.WiththepopularityofAlexNet,in2013,Matthewetal.[10]developedamodeltovisualizeandunderstandtheconvolutionalnetwork,attemptingtooutdothemodeldevelopedbyKrizhevskyetal.AfterthevisualizemodelofKrizhevskyetal.itisobservedthatthesmallchangesinarchitectureimprovedclassificationperformance.TheonlydisadvantageofAlexNetisthatthemodelhadtoomanyparameters.ExtendingourdiscussionaboutCNN,NINetal.[11]teamdevelopedthenetworkwhichutilizedafewernumberofparameters.Themodelhad7.5millionparameters,ascomparedtoAlexNet’s60millionparameters.Furthermore,Googleteampurposedanew,deepCNNmodel,calledtheInception[12].Thismodelreducedthenetworkparametersto4millionascomparedtoAlexNet’s60millionpara