融合显式和隐含语义概念的视频检索技术研究

项目名称： 融合显式和隐含语义概念的视频检索技术研究

项目编号： No.60873165

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 唐胜

作者单位： 中国科学院计算技术研究所

项目金额： 30万元

中文摘要： 在为期三年的项目实施期内，我们严格按照项目计划目标和内容，针对如何建立视频的低层特征和高层语义描述之间的映射，以有效克服所谓的"语义鸿沟"，开展了基于语义概念的视频检索关键技术研究。根据近几年视频检索国际评测TRECVID，确定了语义概念集，通过深入研究如何建立从低层特征到语义概念（语义概念检测）和从用户查询到语义概念（查询分析）的两层映射关系，我们不仅取得了如下预期成果：提出了一种基于密度的自适应隐含语义概念个数确定方法、基于隐含语义分析的语义概念检测方法(LDA-SVM)、基于查询样例概念分布的查询分析方法；而且在视频特征提取、基于跨域迁移学习/多核分类器集成学习/稀疏集成学习的语义概念检测、基于融合隐含语义和显式语义的二分图传播/多模态融合扩展的查询分析与检索方面，我们超预期地提出了一系列新方法。在此基础上，我们按照计划，如期提出了一种基于显性与隐性语义概念融合的三层视频检索方案，构建了基于语义概念的视频检索系统，参加视频检索国际权威评测TRECVID，取得了优异成绩，提高了我们的国际学术影响力。

中文关键词： 视频检索；语义概念检测；隐含语义分析；查询分析

英文摘要： During the past 3 years, we executed the project strictly according to its goal and contents. Aiming at how to establish the mapping between the low-level features and high-level semantic descriptions of video content to bridge up the "semantic gap" efficiently, we investigated the research on the key technologies related to concept-based video retrieval. Based on the concepts selected from those in the recent TREC Video Retrieval Evaluations (TRECVID) and through deep investigations on how to map between the low-level features and high-level descriptions (concept detection) and how to map between user's query and concepts (query analysis), we achieved not only the expected novel results: an adaptive selection of the number of hidden semantic concepts based on density, a concept detection method based on latent semantic analysis, and a query analysis approach by analyzing concept score distributions among within and between relevant and irrelevant sets of query examples; but also series of new approaches (which are out of our expectation) on feature extraction of videos, concept detection based on cross-domain transfer learning / multi-kernel ensemble learning / sparse ensemble learning, and query analysis on explicit and implicit concept-based video retrieval with bipartite graph propagation model / multi-model query expansion. Consequently, as expected by our original plan of project, we proposed a three-layer video retrieval scheme based on fusion of explicit and latent concept detection, and constructed a demonstration system of video retrieval which achieved top ranks in the international authoritative TRECVID evaluations，hence increased our academic impact throughout the world.

英文关键词： Video Retrieval; Semantic Concept Detection; Latent Semantic Analysis; Query Analysis