面向网络多视频摘要的关联学习关键技术研究

项目名称： 面向网络多视频摘要的关联学习关键技术研究

项目编号： No.61472273

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 冀中

作者单位： 天津大学

项目金额： 81万元

中文摘要： 网络多视频摘要主要研究同主题下海量网络视频资源的高效内容分析及整合技术，其中一个关键技术是对视频内容间关联关系的有效分析，即将大而冗余的视频内容空间向小而简约的内容空间有效的转换。而这方面的研究仍处于起步阶段，尚停留在采用传统的表示方法和统计模型进行分析的阶段。针对这一现状，受到近年来关联学习相关成果的启发，本项目将主要从关联推断、传播与排序机制、多样性分析三个方面展开研究。具体包括：1）基于多层贝叶斯和最大信息系数方法的关联推断模型的设计；2）在概率超图模型框架内，研究基于随机游走和流形排序的相关性传播与排序机制；3）基于上述两项研究内容，分别引入半监督关联聚类和吸收性的方法研究多样性分析策略；4）结合互联网中的新闻、体育等类型的视频构建原型系统，验证算法的先进性和实用性。本项目的实施将丰富和完善视频摘要和关联学习领域的基础理论研究，有效推动多媒体浏览、移动互联网等相关产业的发展。

中文关键词： 视频分析；多视频摘要；关联学习；相关性排序；多样性分析

英文摘要： Multi-video abstraction studies the efficient content analysis and integration for large-scale web videos. It is an important research topic in both academic and industrial domains. A critical problem in this research topic is to explore the complex associative relationships among the video contents. However, current research works are still in a fledging stage, in which traditional representation methods and statistical models are employed. Enlightened by the studies on association learning, in this project, we tackle this problem mainly from association inference, relevance ranking, and diversity analysis respectively. Specifically, the following issues would be addressed: 1) The design of associative inference models with Multi-Layer Bayesian and Maximal Information Coefficient methods respectively; 2) The mechanism of novel relevance propagation and ranking algorithms with random walk and manifold ranking respectively in the framework of probabilistic Hypergraph model; 3) Diversity analysis study with semi-supervised association clustering and absorbing ideas respectively on the basis of the above two issues; 4) Constructing a prototype system with news or sports videos to verify the advanced nature and practical value of the proposed methods. The implementation of this research project will enrich and develop the basic theories of both video abstraction and associative learning, as well as promoting the development of related industries, such as multimedia browsing and mobile internet.

英文关键词： Video Analysis;Multi-Video Abstraction;Association Learning;Relevance Ranking;Diversity Analysis