2014-ICLR-OverFeat_ Integrated Recognition, Localization and Detection using Convolutional Networks

《2014-ICLR-OverFeat_ Integrated Recognition, Localization and Detection using Convolutional 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、OverFeat:IntegratedRecognition,LocalizationandDetectionusingConvolutionalNetworksPierreSermanetDavidEigenXiangZhangMichaelMathieuRobFergusYannLeCunCourantInstituteofMathematicalSciences,NewYorkUniversity719Broadway,12thFloor,NewYork,NY10003sermanet,deigen,xiang,mat

2、hieu,fergus,yann@cs.nyu.eduAbstractWepresentanintegratedframeworkforusingConvolutionalNetworksforclassi-fication,localizationanddetection.WeshowhowamultiscaleandslidingwindowapproachcanbeefficientlyimplementedwithinaConvNet.Wealsointroduceanoveldeeplearningapproachto

3、localizationbylearningtopredictobjectbound-aries.Boundingboxesarethenaccumulatedratherthansuppressedinordertoincreasedetectionconfidence.Weshowthatdifferenttaskscanbelearnedsimul-taneouslyusingasinglesharednetwork.Thisintegratedframeworkisthewinnerofthelocalizationt

4、askoftheImageNetLargeScaleVisualRecognitionChallenge2013(ILSVRC2013)andobtainedverycompetitiveresultsforthedetectionandclassificationstasks.Inpost-competitionwork,weestablishanewstateoftheartforthedetectiontask.Finally,wereleaseafeatureextractorfromourbestmodelcalle

5、dOverFeat.1IntroductionRecognizingthecategoryofthedominantobjectinanimageisataskstowhichConvolutionalarXiv:1312.6229v4[cs.CV]24Feb2014Networks(ConvNets)[17]havebeenappliedformanyyears,whethertheobjectswerehandwrittencharacters[16],housenumbers[24],texturelesstoys[1

6、8],trafficsigns[3,26],objectsfromtheCaltech-101dataset[14],orobjectsfromthe1000-categoryImageNetdataset[15].TheaccuracyofConvNetsonsmalldatasetssuchasCaltech-101,whiledecent,hasnotbeenrecord-breaking.However,theadventoflargerdatasetshasenabledConvNetstosignificantlya

7、dvancethestateoftheartondatasetssuchasthe1000-categoryImageNet[5].ThemainadvantageofConvNetsformanysuchtasksisthattheentiresystemistrainedendtoend,fromrawpixelstoultimatecategories,therebyalleviatingtherequirementtomanuallydesignasuitablefeatureextractor.Themaindis

8、advantageistheirravenousappetiteforlabeledtrainingsamples.Themainpointofthispaperistoshowthattrainingaconvolutionalnetworktosimultaneouslyclassif