CopyRight 2012-2014 DS文库版权所有
TernausNet
(0 次评价)1151 人阅读0 次下载

 TernausNet:U-NetwithVGG11EncoderPre-TrainedonImageNetforImageSegmentation VladimirIglovikov LyftInc.SanFrancisco,CA94107,USAEmail:iglovikov@gmail.com AlexeyShvets MassachusettsInstituteofTechnologyCambridge,MA02142,USAEmail:shvets@mit.edu Abstract —Pixel-wiseimagesegmentationisdemandingtaskincomputervision.ClassicalU-Netarchitecturescomposedofen-codersanddecodersareverypopularforsegmentationofmedicalimages,satelliteimagesetc.Typically,neuralnetworkinitializedwithweightsfromanetworkpre-trainedonalargedatasetlikeImageNetshowsbetterperformancethanthosetrainedfromscratchonasmalldataset.Insomepracticalapplications,partic-ularlyinmedicineandtrafficsafety,theaccuracyofthemodelsisofutmostimportance.Inthispaper,wedemonstratehowtheU-Nettypearchitecturecanbeimprovedbytheuseofthepre-trainedencoder.Ourcodeandcorrespondingpre-trainedweightsarepubliclyavailableathttps://github.com/ternaus/TernausNet.Wecomparethreeweightinitializationschemes:LeCununiform,theencoderwithweightsfromVGG11andfullnetworktrainedontheCarvanadataset.Thisnetworkarchitecturewasapartofthewinningsolution(1stoutof735)intheKaggle:CarvanaImageMaskingChallenge. Keywords — ComputerVision,ImageSegmentation,ImageRecognition,Deeplearning,MedicalImageProcessing,SatelliteImagery. I.I NTRODUCTION Recentprogressincomputerhardwarewiththedemoc-ratizationtoperformintensivecalculationshasenabledre-searcherstoworkwithmodels,thathavemillionsoffreeparameters.Convolutionalneuralnetworks(CNN)haveal-readydemonstratedtheirsuccessinimageclassification,objectdetection,sceneunderstandingetc.Foralmostanycomputervisionproblems,CNN-basedapproachesoutperformothertechniquesandinmanycasesevenhumanexpertsinthecor-respondingfield.Nowalmostallcomputervisionapplicationtrytoinvolvedeeplearningtechniquestoimprovetraditionalapproaches.Theyinfluenceoureverydaylivesandthepotentialusesofthesetechnologieslooktrulyimpressive.Reliableimagesegmentationisoneoftheimportanttasksincomputervision.Thisproblemisespeciallyimportantformedicalimagingthatcanpotentiallyimproveourdiagnosticabilitiesandinsceneunderstandingtomakesafeself-drivingvehicles.Denseimagesegmentationessentiallyinvolvesdi-vidingimagesintomeaningfulregions,whichcanbeviewedasapixellevelclassificationtask.Themoststraightforward(andslow)approachtosuchproblemismanualsegmentationoftheimages.However,thisisatime-consumingprocessthatispronetomistakesandinconsistenciesthatareunavoidablewhenhumandatacuratorsareinvolved.Automatingthetreat-mentprovidesasystematicwayofsegmentinganimageontheflyassoonastheimageisacquired.Thisprocessrequires providingnecessaryaccuracytobeusefulintheproductionenvironment.Inthelastyears,differentmethodshavebeenproposedtotackletheproblemofcreatingCNN’sthatcanproduceasegmentationmapforanentireinputimageinasingleforwardpass.Oneofthemostsuccessfulstate-of-the-artdeeplearningmethodisbasedontheFullyConvolutionalNetworks(FCN)[ 2 ].ThemainideaofthisapproachistouseCNNasapowerfulfeatureextractorbyreplacingthefullyconnectedlayersbyconvolutiononetooutputspatialfeaturemapsinsteadofclassificationscores.Thosemapsarefurtherupsampledtoproducedensepixel-wiseoutput.ThismethodallowstrainingCNNintheendtoendmannerforsegmentationwithinputimagesofarbitrarysizes.Moreover,thisapproachachievedanimprovementinsegmentationaccuracyovercommonmethodsonstandarddatasetslikePASCALVOC[ 3 ].ThismethodhasbeenfurtherimprovedandnowknownasU-Netneuralnetwork[ 4 ].TheU-Netarchitectureusesskipconnectionstocombinelow-levelfeaturemapswithhigher-levelones,whichenablesprecisepixel-levellocalization.Alargenumberoffeaturechannelsinupsamplingpartallowspropagatingcontextinformationtohigherresolutionlayers.Thistypeofnetworkarchitectureproventhemselvesinbinaryimagesegmentationcompetitionssuchassatelliteimageanalysis[ 5 ]andmedicalimageanalysis[ 6 ],[ 7 ]andother[ 9 ].Inthispaper,weshowhowtheperformanceofU-Netcanbeeasilyimprovedbyusingpre-trainedweights.Asanexample,weshowtheapplicationofsuchapproachtoAerialImageLabelingDataset[ 8 ],thatcontainsaerospaceimagesofseveralcitieswithhighresolution.Eachpixeloftheimagesislabeledasbelongingtoeither”building”or”not-building”classes.AnotherexampleofthesuccessfulapplicationofsuchanarchitectureandinitializationschemeisKaggleCarvanaimagesegmentationcompetition[ 9 ],whereoneoftheauthorsuseditasapartofthewinning(1stout735teams)solution.II.N ETWORK A RCHITECTURE Ingeneral,aU-Netarchitectureconsistsofacontractingpathtocapturecontextandofasymmetricallyexpandingpaththatenablespreciselocalization(seeforexampleFig. 1 ).Thecontractingpathfollowsthetypicalarchitectureofaconvolutionalnetworkwithalternatingconvolutionandpoolingoperationsandprogressivelydownsamplesfeaturemaps,increasingthenumberoffeaturemapsperlayeratthesametime.Everystepintheexpansivepathconsistsofanupsamplingofthefeaturemapfollowedbyaconvolution. arXiv:1801.05746v1 [cs.CV] 17 Jan 2018

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 14:21 , Processed in 0.862458 second(s), 30 queries .

返回顶部