利用超图随机游走的视频摘要生成方法  

Video Summarization with Random Walk on Hypergraph

作　　者：冀中 樊帅飞 

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072

出　　处：《小型微型计算机系统》2017年第11期2535-2540,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61472273)资助

摘　　要：基于图模型的视频摘要算法在视频摘要技术中有着广泛的应用,通常将视频帧作为图模型的顶点,每两个顶点之间通过一条边相连接.虽然图模型能够很好地捕获每对顶点之间的关系,却不能够较好地捕获多个顶点之间的关系.实际上,视频中往往存在着较为复杂的高阶关系.为此,本文使用超图模型来捕获多个顶点之间的高阶关系,并且基于超图的随机游走算法提出了一种新颖的静态视频摘要方法(Random Walk on Hypergraph for Video Summarization,简称RWH).RWH算法首先通过构建视频概率超图模型,然后利用超图上的随机游走算法选取关键帧.在常用的VSUMM数据集上进行了大量的主观和客观实验,通过与其它视频摘要算法的比较验证了所提RWH算法的先进性.Graph model based video summarization method has been widely used. Typically,it takes video frames as vertices and connects two vertices as edges. Although the graph model can well capture the relationship of vertical pairs,it fails to capture the relationship for more than two vertices. Actually,there usually exist complex high order relationships in a video. To this end,this paper employs the hypergraph model to capture the complex and higher-order relations in a video,and develops a novel method with the random walk on hypergraph to generate a static video summarization. The proposed method is called RWH,which constructs a video probabilistic hypergraph model firstly,then select key frames with the algorithm of random walk on hypergraph. Finally,extensively subjective and objective experiments on the popular VSUM M database indicate the superiority of the proposed RWH algorithm.

关 键 词：随机游走 概率超图 视频摘要 

分 类 号：TP37[自动化与计算机技术—计算机系统结构]