一种基于视频结构的场景分割方法

A Scene Segmentation Approach Based on Video Structure

场景是视频结构中的语义单元。因此基于场景的视频分割,将会对视频的内容分析、检索和浏览提供有益的帮助。提出了一种新的场景分割算法,它利用流形学习理论获得视频的结构特征,然后用马尔科夫链蒙特卡罗方法(Markov chain Monte Carlo,MCMC)动态地进行模拟采样,寻找场景边界的最大后验概率分布,完成场景的自动分割。通过发掘视频结构的内在特征并考虑它的局部信息,使算法能够适合不同类型的视频数据。实验结果也证明了该方法的有效性。

Scene is a semantic unit of video. And it is useful for video analysis, browsing and retrieval. In this paper, we propose a novel scene segmentation method. It consists of two main steps. First, we use the manifold theory to explore the essential feature of video structure. Then Markov chain Monte Carlo （MCMC） algorithm that can find the maxima of the posterior distribution of scene boundaries is adopted to achieve segmentation automatically. Our method has principally focused on underlying video structure and local information that is suitable for various video genres. Experimental results demonstrate the effectiveness of our algorithm.