[NIPS 2011] Matrix Completion for Multi-label Image Classification

《[NIPS 2011] Matrix Completion for Multi-label Image Classification》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、MatrixCompletionforMulti-labelImageClassificationRicardoS.Cabral†,‡FernandoDelaTorre‡JoãoP.Costeira†,AlexandreBernardino†‡CarnegieMellonUniversity,†ISR-InstitutoSuperiorTécnico,Pittsburgh,PALisboa,Portugalrscabral@cmu.edu,ftorre@cs.cmu.edu,{jpc,alex}@

2、isr.ist.utl.ptAbstractRecently,imagecategorizationhasbeenanactiveresearchtopicduetotheurgentneedtoretrieveandbrowsedigitalimagesviasemantickeywords.Thispaperfor-mulatesimagecategorizationasamulti-labelclassificationproblemusingrecentadvancesinmatrixco

3、mpletion.Underthissetting,classificationoftestingdataisposedasaproblemofcompletingunknownlabelentriesonadatamatrixthatconcatenatestrainingandtestingfeatureswithtraininglabels.WeproposetwoconvexalgorithmsformatrixcompletionbasedonaRankMinimizationcrite

4、rionspecificallytailoredtovisualdata,andproveitsconvergenceproperties.Amajoradvantageofourapproachw.r.t.standarddiscriminativeclassificationmethodsforimagecategorizationisitsrobustnesstooutliers,backgroundnoiseandpar-tialocclusionsbothinthefeatureandla

5、belspace.Experimentalvalidationonseveraldatasetsshowshowourmethodoutperformsstate-of-the-artalgorithms,whileeffectivelycapturingsemanticconceptsofclasses.1IntroductionWiththeever-growingamountofdigitalimagedatainmultimediadatabases,thereisagreatneedf

6、oralgorithmsthatcanprovideeffectivesemanticindexing.Categorizingdigitalimagesusingkey-words,however,isthequintessentialexampleofachallengingclassificationproblem.Severalas-pectscontributetothedifficultyoftheimagecategorizationproblem,includingthelargev

7、ariabilityinappearance,illuminationandposeofdifferentobjects.Moreover,inthemulti-labelsettingtheinteractionbetweenobjectsalsoneedstobemodeled.Overthelastdecade,progressintheimageclassificationproblemhasbeenachievedbyusingmorepowerfulclassifiersandbuild

8、ingorlearningbetterimagerepresentations.Ononehand,standarddiscriminativeapproachessuchasSupportVectorMachinesorBoostinghavebeenextendedtothemulti-labelcase[28,14]andincorporatedunderframeworkssuchasMultipleInstanceLearning[31,33,32,20,27]andMulti-tas