基于多证据融合的视频排序方法  

Video Ranking with Multi-Evidence Combination

作　　者：韦世奎[1] 赵耀[1] 朱振峰[1] 

机构地区：[1]北京交通大学信息科学研究所,北京100044

出　　处：《电子学报》2010年第1期167-171,166,共6页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60602030;60776794);国家863高技术研究发展计划(No.2007AA01Z175);国家重点基础研究发展计划(No.2006CB303104);教育部长江学者和创新团队发展计划(No.IRT707);模式识别国家重点实验室开放基金

摘　　要：在视频检索中,通过对用户行为特性的分析发现,用户通常只关注排在最前面的返回结果,而很少有耐心将所有的返回结果浏览一遍.因此,对于一个搜索引擎来说,能否将最相关的结果排在最前面是至关重要的.为了实现这一目标,本文提出了一种基于多证据融合的视频排序方法.该方法利用Dempster-Shafer证据推理理论来协同地融合多方证据,进而推断出最相关的视频镜头.如果多方证据一致,则证明某个视频镜头是相关的,此镜头被认为是最相关的镜头,并被排在返回列表的最前列.相反,如果多方证据产生冲突,那么此镜头就将被排在后面.实验结果表明,利用建议的多证据融合排序算法,搜索引擎的搜索质量,特别是排在前列的搜索结果的准确性,有了明显的改善.According to the analysis on users＇query behavior,we found that the users are rarely patient to go through all the returned results.Usually,they only check the most top returned results,so it is vital for a search engine to rank high relevant results to the top.In order to move vigorously toward this goal,we proposed a video ranking method based on multi-evidence combination.Its theoretical foundation is Dempster-Shafer evidence combination theory,which can fuse the evidence from distinct views and cooperatively infer the most likely relevant results.If the evidence from multiple aspects reaches a consensus on the conclusion that a certain video shot is relevant to query,the video shot is treated as the most likely relevant shot and is ranked to the top of the returned result list.On the contrary,a lack of consensus leads to the judgment that the video shot is not highly relevant to the query and is ranked to the bottom of the returned result list.Experimental results show that the search quality,especially on the most top-ranked results,is significantly improved after using the proposed scheme.

关 键 词：多证据融合 视频排序 检索 DEMPSTER-SHAFER理论 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]