资源描述:
《Contextual Hypergraph Modelling for Salient Object Detection》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、2013IEEEInternationalConferenceonComputerVisionContextualHypergraphModelingforSalientObjectDetectionXiLi,YaoLi,ChunhuaShen,AnthonyDick,AntonvandenHengelAustralianCenterforVisualTechnologies,UniversityofAdelaide,AustraliaAbstractSalientobjectdetectionaimstolocateobjects
2、thatcap-turehumanattentionwithinimages.Previousapproachesoftenposethisasaproblemofimagecontrastanalysis.Inthiswork,wemodelanimageasahypergraphthatuti-lizesasetofhyperedgestocapturethecontextualproper-tiesofimagepixelsorregions.Asaresult,theproblemofsalientobjectdetecti
3、onbecomesoneoffindingsalientver-ticesandhyperedgesinthehypergraph.Themainadvan-tageofhypergraphmodelingisthatittakesintoaccounteachpixel’s(orregion’s)affinitywithitsneighborhoodasImageSVMsaliencyHypergraphsaliencywellasitsseparationfromimagebackground.Further-Figure1:Ill
4、ustrationofourapproachestosalientobjectdetection.more,weproposeanalternativeapproachbasedoncenter-versus-surroundcontextualcontrastanalysis,whichper-versus-surroundfeaturediscrepancy[6,8–10,13–15].formssalientobjectdetectionbyoptimizingacost-sensitiveGlobalsalientobjec
5、tdetectionapproaches[4,5,7,11,12]supportvectormachine(SVM)objectivefunction.Experi-estimatethesaliencyofaparticularimageregionbymea-mentalresultsonfourchallengingdatasetsdemonstratethesuringitsuniquenessintheentireimage.Theseap-effectivenessoftheproposedapproachesagain
6、stthestate-proachesmodeluniquenessbyexploitingtheglobalsta-of-the-artapproachestosalientobjectdetection.tisticalpropertiesoftheimage,includingfrequencyspec-trumanalysis[4],color-spatialdistributionmodeling[7],high-dimensionalGaussianfiltering[11],low-rankmatrix1.Introdu
7、ctiondecomposition[12],andgeodesicdistancecomputation[5].Therefore,thedefinitionofobjectsaliencydependsonImagesaliencydetectionaimstoeffectivelyidentifythechoiceofcontext.Globalsaliencydefinesthecontextimportantandinformativeregionsinimages.Earlyap-astheentireimage,where
8、aslocalsaliencyrequiresthedef-proachesinthisareafocusmainlyonpredictingwhereinitionofalocalcontext.Inthiswork,wefirsts