资源描述:
《Correlation-based Deep Learning for Multimedia Semantic Concept Detection基于相关深度学习分享检测》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、Correlation-basedDeepLearningforMultimediaSemanticConceptDetectionHsin-YuHa,YiminYang,SamiraPouyanfar,HaimanTian,andShu-ChingChenSchoolofComputingandInformationSciences,FloridaInternationalUniversity,Mami,FL33199,USA{hha001,yyang010,spouy001,htian005,chens}@
2、cs.fiu.eduAbstract.Nowadays,conceptdetectionfrommultimediadataiscon-sideredasanemergingtopicduetoitsapplicabilitytovariousapplica-tionsinbothacademiaandindustry.However,therearesomeinevitablechallengesincludingthehighvolumeandvarietyofmultimediadataaswellasi
3、tsskeweddistribution.Tocopewiththesechallenges,inthispaper,anovelframeworkisproposedtointegratetwocorrelation-basedmethods,Feature-CorrelationMaximumSpanningTree(FC-MST)andNegative-basedSampling(NS),withawell-knowndeeplearningalgo-rithmcalledConvolutionalNeu
4、ralNetwork(CNN).First,FC-MSTisintroducedtoselectthemostrelevantlow-levelfeatures,whichareex-tractedfrommultiplemodalities,andtodecidetheinputlayerdimensionoftheCNN.Second,NSisadoptedtoimprovethebatchsamplingintheCNN.UsingNUS-WIDEimagedatasetasaweb-basedappli
5、ca-tion,theexperimentalresultsdemonstratetheeectivenessofthepro-posedframeworkforsemanticconceptdetection,comparingtootherwell-knownclassiers.Keywords:DeepLearning,FeatureSelection,Sampling,SemanticCon-ceptDetection,Web-basedMultimediaData1IntroductionInre
6、centdecades,thenumberofmultimediadatatransferredviatheInternetincreasesrapidlyineveryminute.Multimediadata,whichreferstodatacon-sistingofvariousmediatypesliketext,audio,video,aswellasanimation,isrichinsemantics.Tobridgethesemanticgapbetweenthelow-levelfeatur
7、esandhigh-levelconcepts,itintroducesseveralinterestingresearchtopicslike,datarepresentations,modelfusion,imbalanceddataissue,reductionoffeaturedimensions,etc.Becauseoftheexplosivegrowthofmultimediadata,thecomplexityrisesexponentiallywithlinearlyincreasingdim
8、ensionsofthedata,whichposesagreatchallengetomultimediadataanalysis,especiallysemanticconceptdetection.Duetothisfact,itdrawsmultimediasociety'sattentiontoidentifyusefulfeaturesubsets,reducethefea