资源描述:
《convolutional networks can learn to generate affinity graphs for image segmentationnew》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、LETTERCommunicatedbyDmitriChklovskiiConvolutionalNetworksCanLearntoGenerateAffinityGraphsforImageSegmentationSrinivasC.Turaga∗sturaga@mit.eduDepartmentofBrainandCognitiveSciences,MassachusettsInstituteofTechnology,Cambridge,MA02139,U.S.A.JosephF.Murray∗jfmu
2、rray@jfmurray.orgDepartmentofBrainandCognitiveSciences,MassachusettsInstituteofTechnology,HowardHughesMedicalInstitute,Cambridge,MA02139,U.S.A.VirenJainviren@mit.eduDepartmentofBrainandCognitiveSciences,MassachusettsInstituteofTechnology,Cambridge,MA02139,
3、U.S.A.FabianRothfabian.roth@gmail.comDepartmentofBrainandCognitiveSciences,MassachusettsInstituteofTechnology,HowardHughesMedicalInstitute,Cambridge,MA02139,U.S.A.MoritzHelmstaedterMoritz.Helmstaedter@mpimf-heidelberg.mpg.deKevinBriggmanKevin.Briggman@mpim
4、f-heidelberg.mpg.deWinfriedDenkWinfried.Denk@mpimf-heidelberg.mpg.deMax-PlanckInstituteforMedicalResearch,D-69120,Heidelberg,GermanyH.SebastianSeungseung@mit.eduDepartmentsofBrainandCognitiveSciencesandPhysics,MassachusettsInstituteofTechnology,HowardHughe
5、sMedicalInstitute,Cambridge,MA02139,U.S.A.Manyimagesegmentationalgorithmsfirstgenerateanaffinitygraphandthenpartitionit.Wepresentamachinelearningapproachtocomputing∗Theseauthorscontributedequallytothewritingofthisletter.NeuralComputation22,511–538(2010)C200
6、9MassachusettsInstituteofTechnology512S.Turagaetal.anaffinitygraphusingaconvolutionalnetwork(CN)trainedusinggroundtruthprovidedbyhumanexperts.TheCNaffinitygraphcanbepairedwithanystandardpartitioningalgorithmandimprovessegmenta-tionaccuracysignificantlycompare
7、dtostandardhand-designedaffinityfunctions.Weapplyouralgorithmtothechallenging3Dsegmentationproblemofreconstructingneuronalprocessesfromvolumetricelectronmicroscopy(EM)andshowthatweareabletolearnagoodaffinitygraphdirectlyfromtherawEMimages.Further,weshowthato
8、uraffinitygraphim-provesthesegmentationaccuracyofbothsimpleandsophisticatedgraphpartitioningalgorithms.Incontrasttopreviouswork,wedonotrelyonpriorknowledgeintheformofhand-designedimagefeaturesorimagepreprocess