深度学习在图像识别中的分析及应用-analysis and application of depth learning in image recognition.docxVIP

AbstractPreciserecognitionforimageHasveryimportantresearchsignificance,imagerecognitiontechnologyiswidelyusedinMedicine,spacemilitary,industryandagriculture.AsnowmostmethodofimagerecognitionUsedartificialfeatureextractionwhichNotonlylaborious,butalsodifficulttoextract.DeepLearningisakindofunsupervisedlearning,Inthelearningprocessweneednotknowthevaluesofsamples,Thewholeprocesscanalsoextractedgoodcharacteristicswithouthumanparticipation.inrecentyears,Thedeeplearningusedinimagerecognitionbecomethehotresearchtopicinthefieldofimagerecognition,Hasachievedgoodeffect,andhaveabroadspaceforresearch.Inthispaper,webasedonthestudyonthetheoryofimagerecognitionanalyzesthedeeplearningthebasicmodelsandmethods,thendoexperimentonsomeimagedatasets.Givendeeplearningmoreforlargesampleset,weimprovedaalgorithmproposedtouseitintosmallsampleset,theworkcanbedescribedasfollows:(1)Analysisbasicprinciplesoftheconvolutionneuralnetworks(CNNs),introducethetrainingprocessandmodelstructureofit.Theconvolutionlayercanmaketheoriginalsignalenhancement,andreducenoiseaswellasimprovesignal-to-noiseratiobyconvolutionoperation,usethemodelintothehandwritingdatasetMNIST,comparedtootherclassicalalgorithms,analyzetheiradvantagesanddisadvantagesabouttimeandrecognitionrate.(2)aimattheinadequaciesoftheconvolutionneuralnetwork,Analysisthebasicprinciples,trainingprocessandmodelstructureofdeepbeliefnetworks(DBNs).ThestratifiedtrainingmechanismofDBNsgreatlyreducesthedifficultyandreducesthetrainingtimeofit.WeuseSoftmaxclassificationsystemastheclassifier,theusethismodeldoexperimentonMNISTdatasets,comparedtoconvolutionneuralnetwork,mainlyintherecognitionrateandtime-consuming,whichcanbeprovedthatDBNshasthesameidentifyingrateofCNNs,buttheelapsedtimeisgreatlyreduced,andthenanalyzethereasons;addition,usethemodelintotheCIFAR-10databases,comparedtotheotheralgorithms.(3)aimatthedeepbeliefnetworksalgorithmisonlyapplicabletolargedatasets,raiseaimprovedalgorithmofdeepbeliefnetworkaimatsmallsampleset.Beforethepre-training,down-sampli