资源描述：
《ACM-MM2009-Shao》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。
1、Semi-SupervisedTopicModelingforImageAnnotationYuanlongShao,YuanZhou,XiaofeiHe,DengCai,HujunBao{shaoyuanlong,zhouyuan,xiaofeihe,dengcai,bao}@cad.zju.edu.cnStateKeyLaboratoryofCAD&CG,ZhejiangUniversityNo.38,ZhedaRoad,Hangzhou,Zhejiang,P.R.ChinaABSTRACTsearchandaccessimagesthatarewan
2、tedeffectively.Im-ageannotation,thetaskofassociatingtexttothesemanticWeproposeanoveltechniqueforsemi-supervisedimagean-contentofimages,isagoodwaytoreducethesemanticgapnotationwhichintroducesaharmonicregularizerbasedonandcanbeusedasanintermediatesteptoimageretrieval.thegraphLaplacia
3、nofthedataintotheprobabilisticseman-Itenablesuserstoretrieveimagesbytextqueriesandof-ticmodelforlearninglatenttopicsoftheimages.Byusingtenprovidessemanticallybetterresultsthancontent-basedaprobabilisticsemanticmodel,weconnectvisualfeaturesimageretrieval.Inrecentyears,itisobservedt
4、hatimageandtextualannotationsofimagesbytheirlatenttopics.annotationhasattractedmoreandmoreresearchinterests.Meanwhile,weincorporatethemanifoldassumptionintotheThefundamentalproblemofimageannotationishowtomodeltosaythattheprobabilitiesoflatenttopicsofimagesmodeltherelationshipamong
5、differentmodalities,includingaredrawnfromamanifold,sothatforimagessharingsimi-visualfeaturesandtextualannotations,associatedwiththelarvisualfeaturesorthesameannotations,theirprobabilitypossiblyexistedlatenttopicsofimages,aswellastherela-distributionoflatenttopicsshouldalsobesimilar
6、.Wecreatetionshipamongdifferentimages.Latenttopicmodelinghasanearestneighborgraphtomodelthemanifoldandproposelongbeenapromisingapproachforthisproblem[3],[1],[8],aregularizedEMalgorithmtosimultaneouslylearnagen-[9].Asiscommon,modelbasedapproacheshavethebenefiterativemodelandassignpro
7、babilitydensityoflatenttopicsofbetterefficiencyandstability,whileitsuffersmostlyfromtoimagesdiscriminatively.Inthisway,databaseswithveryprobablyinsufficientmodeling,i.e.,whenthemodeldoesfewlabeledimagescanbeannotatedbetterthanpreviousnotfullydescribetheproblemdomain,theinferredquan-wor
8、ks.titiesmaynotbeaccurate,e.g.,if
