CopyRight 2012-2014 DS文库版权所有
Systematic evaluation of CNN advances on the ImageNet
(0 次评价)2602 人阅读0 次下载

 SystematicevaluationofCNNadvancesontheImageNet DmytroMishkin a ,NikolaySergievskiy b ,JiriMatas a a CenterforMachinePerception,FacultyofElectricalEngineering,CzechTechnicalUniversityinPrague.Karlovonamesti,13.Prague2,12135 b ELVEESNeoTek,Proyezd4922,4build.2,Zelenograd,Moscow,RussianFederation,124498 Abstract ThepapersystematicallystudiestheimpactofarangeofrecentadvancesinCNNarchitecturesandlearningmethodsontheobjectcategorization(ILSVRC)problem.Theevalutionteststheinfluenceofthefollowingchoicesofthearchi-tecture:non-linearity(ReLU,ELU,maxout,compatabilitywithbatchnormal-ization),poolingvariants(stochastic,max,average,mixed),networkwidth,classifierdesign(convolutional,fully-connected,SPP),imagepre-processing,andoflearningparameters:learningrate,batchsize,cleanlinessofthedata,etc.Theperformancegainsoftheproposedmodificationsarefirsttestedindivid-uallyandthenincombination.Thesumofindividualgainsisbiggerthantheobservedimprovementwhenallmodificationsareintroduced,butthe”deficit”issmallsuggestingindependenceoftheirbenefits.Weshowthattheuseof128x128pixelimagesissufficienttomakequalitativeconclusionsaboutoptimalnetworkstructurethatholdforthefullsizeCaffeandVGGnets.Theresultsareobtainedanorderofmagnitudefasterthanwiththestandard224pixelimages. Keywords: CNN,benchmark,non-linearity,pooling,ImageNet 1.Introduction Deepconvolutionnetworkshavebecomethemainstreammethodforsolv-ingvariouscomputervisiontasks,suchasimageclassification[1],objectde-tection[1,2],semanticsegmentation[3],imageretrieval[4],tracking[5],textdetection[6],stereomatching[7],andmanyother.Besidestwoclassicworksontrainingneuralnetworks–[8]and[9],whicharestillhighlyrelevant,thereisverylittleguidanceortheoryontheplethoraofdesignchoicesandhyper-parametersettingsofCNNswiththeconsequentthatresearchersproceedbytrial-and-errorexperimentationandarchitecturecopying,stickingtoestablishednettypes.WithgoodresultsinImageNetcompetition, PreprintsubmittedtoComputerVisionandImageUnderstandingJune14,2016 arXiv:1606.02228v2 [cs.NE] 13 Jun 2016

打分:

0 星

用户评论:

小飞飞
于 2020-07-15 上传

版权及免责声明|RISC-V单片机中文网 |网站地图

GMT+8, 2022-10-2 14:01 , Processed in 0.554331 second(s), 30 queries .

返回顶部