基于多模态概念关联图的视频检索  

An Effective Video Retrieval Approach Based on Multi-modality Concept Correlation Graph

作　　者：冯柏岚[1,2] 包蕾[1,2] 曹娟[1] 张勇东[1] 林守勋[1] 

机构地区：[1]中国科学院计算技术研究所前瞻研究实验室,北京100190 [2]中国科学院研究生院,北京100049

出　　处：《计算机辅助设计与图形学学报》2010年第5期827-832,共6页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家"九七三"重点基础研究发展计划项目(2007CB311100);国家"八六三"高技术研究发展计划(2007AA01Z416);国家自然科学基金(60902090;60802067);北京市教育委员会共建项目专项资助

摘　　要：为了有效地提高基于概念的视频检索的检索性能,提出一种新颖的基于多模态概念关联图的视频检索方法.首先通过分析查询与概念之间的组织关系得到网状关系模型描述,并基于该模型构建概念关联图;然后提出查询与概念的多模态映射结构,将多模态查询融入概念关联图,增强概念扩展的针对性;之后使用流形排序动态地扩展索引概念集;全局稳态后采用正交的概念融合方法计算视频索引值,用于视频检索.与多种典型的基于概念的视频检索方法相比,文中方法的平均检索精度增幅达14.6%～86.2%.此外,实验结果表明,该方法在实际的交互式视频检索系统中也具有良好的适用性.A novel multi-modality concept correlation graph approach is proposed to promote the performance of concept-based video retrieval.Firstly,a net-style concept correlation graph is constructed using the original similarities among the concepts.Then,query is added into the graph based on a multi-modality mapping strategy between query and concept.Thirdly,ranking on manifolds algorithm is used to dynamically diffuse the concept correlations in this graph until a global stable state is achieved.Finally,the selected concepts are orthogonally fused to implement the video retrieval.In the experiments of automatic video retrieval,our approach has achieved great improvement from 14.6% to 86.2% over the state-of-art methods.Moreover,we also apply it to our interactive retrieval system and gains excellent performance and user experience.

关 键 词：基于概念的视频检索 概念关联图 流形排序 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]