1、復旦大學碩士學位論文基于高維索引結構的視頻片斷檢索算法研究姓名：劉芳潔申請學位級別：碩士專業(yè)：計算機應用技術指導教師：薛向陽200405152ABSTRACTWiththedevelopmentofnetworkandmultimediatechniques，videoprocessingandreuievalhasbecomeallactiveresearchfieldTheresearchtasksincludekeyframeex

2、tractionsimilaritycomputationbetweenvideoclipsand11ighdimensionalindexstrucRkresetcInthisthesis，wefocusonhowtealizethevideoclipretrievalwithefficiencyanddiscusssomerelatedtechniques，includinghighdimensionalindexstructttr

3、es，similaritymodelaboutvideoclipsandsomevideoclipretrievalalgorithmsbasedondifferentindexstructuresAfteranalyzingthestateofartofexistingindexingstructureandretrievalalgodthmsrelatedtovideoelipretrieval，afastvideoclipretr

4、ievalalgodthrnbasedonVAFileisproposed。VAFilerepresentsavectorapproximatelybycompressingtheoriginaldataanditmaintainstheoriginalorderwhenrepresentingvectorsinasequence，whichisaveryvaluablemeritforvectorsequencesmatchingIn

5、tegratedwitllanewvideoclipsimilaritymodelweproposedthisalgorithmcansupportvideoclipretrievalwithefficiencyBasedonVA—File，anewhighdimensionalindexstructure，namedOrderedVAFile(OVAFile)，Waspresentedinthispaperwhichstoresthe

6、approximatevectorsinto‘proper’positionsoffileTheword‘proper’meanstheapproximatevectorsclosetoeachotherinmetricspacesshouldhaveclosepositionsinvectorapproximationfileByuseofthei／earnei【ghboringmeritsofOVAFile，weproposeane

7、wmethodtocomputethesimilaritybetweentwokey臺alrlesandtwovideoclipsInordertospeedupthequeryprocess，thewholeorderedapproximatevectorfilearesegmentedintomanyconsecutiveslices，ofwhichthecentersandstartendpositionsarerecordeda

8、ccordinglyOnlyasmallnumberofslices，whichhavethehighestprobabilitycontmningthequeryresults，havetobescannedItmeansgreatreductionofdiskIOcostThisapproachisbothevaluatedintheoryandexperimentsTheexperimentalresultsshowthatoar

9、newalgorithmincrediblyshortenedtheretrievaltimeandCanbeappliedeffectivelytomanysimilarityretrievalbasedonhighdimensionalvectorsequencesKeywords：VideoClipRetrieval，VideoClipSimilarityModel，Hi曲DimensionalIndex，VA—File，OVA—