视频分割和关键帧提取统一框架

ISSN100020054CN1122223öN()JTsinghuaUniv(Sci&Tech),200545S12005,Vol.45,No.S124ö391838218411,2,3,2,3,3(1.,100070;2.,100876;3.,066004):2005205220:(60472082);(XWSP):(19632),(),,E2mail:[e-mail]:,,,,,SVM,21h(10250),:;;:TP391:A:100020054(2005)S121838204Frameworkforshotboundarydetectionandkey-frameextractionFENGHuamin1,FANGWei2,LIUSen3,FANGYong2,3,SONGGuosen3(1.KeyLaboratoryforSecurityandSecrecyofInformation,BeijingElectronicScienceandTechnologyInstitute,Beijing100070,China;2.TelecommunicationEngineeringCollege,BeijingUniversityofPostsandTelecommunications,Beijing100876,China;3.TheCollegeofInformationScienceandEngineering,YanshanUniversity,Qinhuangdao066004,China)Abstract:Shotboundarydetectionisanimportantstepinvideoapplications.Withtherapiddevelopmentofthevideoeditingtechnology,especiallytheextensiveuseofsub2windowinnews,theoldmethodsofvideosegmentationcannotefficientlydetecttheshotboundarycausedbyspecialvideotechnique.Inthispaper,previoustemporalmulti2resolutionanalysis(TMRA)workwasextendedbyadoptingblockedchromatichistogram(BCH)asfeature,usingsupportvectormachines(SVM)toclassifythevideoframeswithinaslidingwindowintonormalframes,gradualtransitionframes,andCUTframes,andclusteringtheframesintodifferentshotcategories.Theresultontruthbed,whichhasabout21h(10250shots)newsvideoclip,showsthatthenewmethodhasrelativelygoodprecisionandrecallfortheshotboundarydetection.Itbasicallyresolvesthedifficultiesofshotboundariesdetectioncausedbysub2windowtechniqueinvideo.Theframeworkalsogreatlyimprovesaccuracyofgradualtransitionsofshot.Keywords:temporalmulti2resolutionanalysis(TMRA);blockedchromatichistogram(BCH);supportvectormachines(SVM),[1],(CUT)(GT),,Hampapur[2],,,Lin[3](temporalmulti2resolutionanalysis,TMRA),,,CUTGTFeng[4],,GTAdcock[5]k(knearestneighbor),,,TMRA[3],,,(supportvectormachines,SVM)1SVM1,5,,,,ö12SVM2.1,[1,3,6],YUV64bin,YUV4÷1÷1,,12(÷=4÷3),,bin,13(BCH),13:VH=[VH1,VH2,,VH13]2.22.2.1,,Gauss,[7]Gauss,,,,,,2.2.2,N13:MH=V1H1V1H2V1H13V2H1V2H2V2H3wVNH1VNH2VNH13.(1),N13:MW=V1W1V1W2V1W13V2W1V2W2V2W3wVNW1VNW2VNW13.(2),,,,,,;,,[4],305,SVM,42.2.3,,260,10,,9381,:11,,143,,143(temporal2windowvector,TWV)MW=wV(N-5)W1V(N-5)W2V(N-5)W13wV(N-1)W1V(N-1)W2V(N-1)W13VNW1VNW2VNW13V(N+1)W1V(N+1)W2V(N+1)W13wV(N+5)W1V(N+5)W2V(N+5)W13w,TWV=[V(N-5)W1V(N-5)W13VNW1VNW2VNW13V(N+5)W13].(3)2.3SVM2.3.1SVM,,SVM2090,[8],SVM,C,k2(k2foldcross2validation),2.3.2SVM,(CF)(GF),(NF),123,3,3SVM,(oneagainstone)[8],33,,SVM,,,22SVM2.4,,,,,[4],2.53:1),;2),;3),,,3CCTVF121h,10250,91.5%90.4%,0481()2005,45(S1),1,TMRA[3]F11TMRATMRA%%F1TMRA2SVM%%F1CUT77.098.00.8793.195.40.94GT65.297.10.7882.780.10.8171.597.00.8291.590.40.91SVM,,,,452,439,97.1%TMRA12.0%,4,,,SVMNFCFGF,21h10250,,91.5%90.4%,,,Web,Web(References)[1]ZHANGHongjiang,KankanhalliA,SmoliarSW.Automaticpartitioningoffull2motionvideo[J].ACMMultimediaSystems,1993,1(1):1028.[2]HampapurA,JainR,WeymouthTE.Digitalvideosegmentation[A].ACMMultimediaConference[C].SanFrancisco,California,UnitedStates,1994.357364.[3]LinY,KankanhalliMS,ChuaTS.Temporalmulti2resolutionanalysisforvideosegmentation[A].ProceedingsofSPIEConferenceStorageandRetrievalforMediaDatabaseVIII,SPIE[C].Fujisawa,Japan,2000.119124.[4]FENGHuamin,ChandrashekharaA,ChuaTS.Anautomatictemporalmulti2resolutionanalysisframeworkforshotboundarydetection[A].MMM2003,the9thInternationalConferenceonMulti2MediaModeling[C].Taipei,China,2003.224.[5]AdcockJ,GirgensohnA,CooperM.FXPALexperimentsforTRECVID2004[A].ProceedingsofTRECVID2004[C].CA,USA,2005.[6]GONGYihong.Anaccuracyandrobustmethodfordetectingvideoshotboundaries[A].ICMCS[C].Florence,Italy,1999.1:850854.[7]WANGYuping,LeeSL.Scale2spacederivedfromB2spline[J].PAMI,1998,20(10):10401055.[8],.[M].:,2004.DENGNaiyang,TIANYingjie.ANewMethodinData2MiningSupportedVectorMachine[M].Beijing:SciencePress,2004.(inChinese)1481,: