Very Deep Convolutional Networks for Large-Scale Image Recognition.pdf

《Very Deep Convolutional Networks for Large-Scale Image Recognition.pdf》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、VeryDeepConvolutionalNetworksforLarge-ScaleImageRecognitionKarenSimonyanAndrewZissermanVisualGeometryGroup,UniversityofOxford{karen,az}@robots.ox.ac.ukAbstractInthisworkweinvestigatetheeffectoftheconvolutionalnetworkdepthonitsaccuracyinthelarge-scaleimagerecogni

2、tionsetting.Ourmaincontributionisathoroughevaluationofnetworksofincreasingdepth,whichshowsthatasignifi-cantimprovementontheprior-artconfigurationscanbeachievedbypushingthedepthto16–19weightlayers.ThesefindingswerethebasisofourImageNetChallenge2014submission,whereou

3、rteamsecuredthefirstandthesecondplacesinthelocalisationandclassificationtracksrespectively.1IntroductionConvolutionalnetworks(ConvNets)haverecentlyenjoyedagreatsuccessinlarge-scalevisualrecognition[10,16,17,19]whichhasbecomepossibleduetothelargepublicimagereposito

4、ries,suchasImageNet[4],andhigh-performancecomputingsystems,suchasGPUsorlarge-scaledis-tributedclusters[3].Inparticular,animportantroleintheadvanceofdeepvisualrecognitionarchi-tectureshasbeenplayedbytheImageNetLarge-ScaleVisualRecognitionChallenge(ILSVRC)[1],whic

5、hhasservedasatestbedforafewgenerationsoflarge-scaleimageclassificationsystems,fromhigh-dimensionalshallowfeatureencodings[13](thewinnerofILSVRC-2011)todeepCon-vNets[10](thewinnerofILSVRC-2012).WithConvNetsbecomingmoreofacommodityinthecomputervisionfield,anumberofa

6、ttemptshavebeenmadetoimprovetheoriginalarchitectureof[10]inabidtoachievebetteraccuracy.Forinstance,thebest-performingsubmissionstotheILSVRC-2013[16,19]utilisedsmallerreceptivewindowsizeandsmallerstrideofthefirstconvolutionallayer.Anotherlineofimprovementsdealtwit

7、htrainingandtestingthenetworksdenselyoverthewholeimageandovermultiplescales[7,16].arXiv:1409.1556v2[cs.CV]15Sep2014Inthispaper,weaddressanotherimportantaspectofConvNetarchitecturedesign–itsdepth.Tothisend,wefixotherparametersofthearchitecture,andsteadilyincreaset

8、hedepthofthenetworkbyaddingmoreconvolutionallayers.Therestofthepaperisorganisedasfollows.InSect.2,wedescribeourConvNetconfigurations.Thedetailsoftheimageclassificationt