基于流形学习和混合模型的视频摘要方法

Video Abstraction Based on Manifold Learning and Mixture Model

视频摘要是进行视频浏览、视频检索、视频索引等视频应用的前提,而且视频摘要类似于文本的摘要,也是对视频内容的一个简短概括。为了自动获得既包含视频的主要信息,而冗余信息又少的视频摘要,提出了一种基于流形学习和有限混合模型的自动视频摘要方法。该方法通过对视频序列进行流形建模,首先得到视频场景的初次分割;然后对包含内容较多的场景,使用等距降维方法计算视频帧的特征向量;最后将视频帧的特征向量输入到混合模型进行聚类分析,得到更细粒度的摘要结果。为了实现视频摘要的自动处理,所采用的混合模型需要具有模型选择功能。混合模型的聚类结果和流形建模的结果共同构成了视频摘要。视频分割片段的实验结果表明,在不需人为干预的情况下,所提供的视频摘要不仅包含视频主要内容,而且冗余信息少。

Techniques for video abstraction has attracted tremendous attention for its application in video browsing, video indexing ,video retrieval and so on. Video abstraction is brief summary of the video content like the text abstraction. In the paper, an automatic method for video abstraction is presented which is based on manifold modeling and mixture model. Man- ifold modeling is applied to generate the scene manifold of the video, Isomap is used to reduce the dimension of the video frames in larger scenes and the 10w dimension vectors are put into the mixture model with model selection to complete clus- ter analysis. Because mixture model with model selection can adapt to the data from any distribution, it is applied to gener- ate the video abstraction automatically. The results from manifold modeling together with those from mixture model constitute the abstraction results. The experiments present the abstraction with less redundance,which demonstrates the effective and efficiency of the proposed method.