Traffic Sign Recognition With Hinge Loss Trained Convolutional Neural Networks

《Traffic Sign Recognition With Hinge Loss Trained Convolutional Neural Networks》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、IEEETRANSACTIONSONINTELLIGENTTRANSPORTATIONSYSTEMS,VOL.15,NO.5,OCTOBER20141991TrafficSignRecognitionWithHingeLossTrainedConvolutionalNeuralNetworksJunqiJin,KunFu,andChangshuiZhang,Member,IEEEAbstractTrafficsignrecognition(TSR)isanimportantandchallengingtaskforintelligenttransportationsy

2、stems.Wede-scribethedetailsofourmodelsarchitectureforTSRandsug-gestahingelossstochasticgradientdescent(HLSGD)methodtotrainconvolutionalneuralnetworks(CNNs).OurCNNconsistsofthreestages(70110180)with1162284trainableparameters.TheHLSGDisevaluatedontheGermanTrafficSignRecognitionBenchmark,w

3、hichoffersafasterandmorestableconvergenceandastate-of-the-artrecognitionrateof99.65%.WewriteagraphicsprocessingunitpackagetotrainseveralCNNsandes-tablishthefinalclassifierinanensembleway.IndexTermsConvolutionalneuralnetworks(CNNs),hingeFig.1.Structureofalayerand3-Dimageofamap.loss,stocha

4、sticgradientdescent(SGD),trafficsignrecognition(TSR).Themodelaccomplishesfeatureextractingandclassifyingasawhole,andbothbuildingblocksarelearnedinasupervisedI.INTRODUCTIONprocedure.ThelearnedfeaturesareusuallyveryrobusttotheRAFFICsignsplayveryimportantrolesforbothtrans-specifictask.Inrec

5、entyears,theCNNhasachievedseveralTportationefficiencyandsafety.Anautomatictrafficsignstate-of-the-artperformancessuchasintheImageNetChal-recognition(TSR)systemishelpfulforassistingdriversandislenge[13]andthe2011InternationalJointConferenceonessentialforautonomouscars.Theresearcharoundthi

6、sissueNeuralNetworks(IJCNN)competition[14],[15].haslongbeenpopular[1],butthetaskischallenging.SomeTheCNNusuallyhasmillionsofparameters,whichismorepopularmethodsincludeBayesianclassifiers[2],boosting[3],difficulttotrainthanflatmodels.Stochasticgradientdescenttreeclassifiers[4],andsupportvec

7、tormachines(SVMs)[5].(SGD)ispreferredfortrainingtheCNN,butaccordingtotheThesemethods,fromtoday’spointofview,areconsideredreportsbySermanetandLeCun[16],Ciresanetal.[14],andusinghand-codedfeaturessuchasacircledetectorin[2],aLeCunetal.[12],SGDtrainingalsorequiresagooddealofHaarwavelet[6