资源描述:
《互联网视频内容检索的关键技术研究》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、中国科学技术大学博士学位论文互联网视频内容检索的关键技术研究姓名:刘守群申请学位级别:博士专业:网络传播系统与控制指导教师:朱明20100501摘要4.提出了基于Map/Reduce的分布式海量视频片段相似度计算方法。视频数据的特征复杂使得视频相似度度量算法在处理大规模视频数据时存在性能瓶颈,针对这一问题,本文首先提出了一种面向海量视频相似度计算的视频特征表示方法,该方法借鉴了信息论的模型以及文本数据相似度的计算方法,在特征描述的基础上,提出了基于Map/Reduce架构的视频相似度计算方法,将视频相似度计算的任务进行分解,并通过分布式的框架进行处理,有效地提高了视频相似度的计算效率,
2、并且具有较高的可扩展性。关键词:互联网视频检索,视频文本索引,视频信息检索,聚类,信息抽取,视频消重,视频数据处理,海量视频数据处理IIAbstractABSTRACTWiththefastdevelopmentofcomputernetworksanddigitalmultimediatechnology,Internetvideoapplicationbecomespopular,theamountofvideoclipsovertheInternetisveryhuge,andhowtoeffectivelydiscovervideocontents,retrievalvideo
3、clipsandprocessingvideodatabecomesachallengingtaskinbothresearchandindustry.State-of-artInternetvideoretrievalsystemsarebasedtextsuchasthetitleandsurrondingdescriptiontextofvideo.Althoughitisfast,therearemanyproblemsinit,suchastheincompletionofindexing,theincorrectofdescriptiontext,whichleedtoth
4、epoorresultinttheretrievalresult.Thesesystemsalsoignorethecontentofvideos,whichleedtomanyduplicationsintheretrievalresult.Accordingtotheseproblems,thispaperfocusesandstudiesthekeyissuesinInternetvideoretrievalsystem,andresearchonthedataminingmethodsofvideocontentsinordertoimprovetheresult.Themai
5、ncontributionsareillustratedasfollows:1.Weproposeasemanticstructurebasedinternetvideodiscoveryandinformationmethod.Basedonthetextmethodofvideoretrieval,westudythediscoverymethodforinternetvideocontainerwebpagesandthevideodescriptiontextextractionalgorithm.First,webpagesarecovertintosemanticstruc
6、turalrepresentation,thenweproposeatreebasedsimilarityalgorithmtoidenfityvideocontainerwebpageandextractvideodescriptiontext.2.WeproposethemodelofvideocontentsimilaritybasedonSIFTfeatureandtimeseriesfeature.First,weproposeaframematchingstrategybasedonSIFTfeature,tospeedupthematchingprocess,LSHhas
7、hingisemployed,thematchingprocsionisimprovedbythefeature.ThenweuseRANSACtofilterthematchingsequenceandeffectivelyremovetheoutliers,whichcombiesthetimeseriesfeature.Themethodimprovestheprecisionofsimilaritymeasurementandiseff