基于SVD和稀疏子空间聚类的视频摘要

发布时间：2019-02-22 11:28

【摘要】：为方便视频的浏览和存储,以概括视频内容为目的的视频摘要技术变得十分重要.针对目前在视频摘要问题中,根据先验知识事先确定和阈值调整2种关键帧数目的确定方法的灵活性及直观性不足这一问题,提出一种基于SVD和稀疏子空间聚类的视频摘要算法.该算法利用SVD对视频在时间维度上进行维数缩减,以累计贡献率为度量标准确定视频在时间维度上的主成分分量个数,将主成分分量个数作为关键帧数目;然后利用稀疏子空间聚类算法对视频帧进行聚类;最后在每一类中选取与其他视频帧相关性最大的帧作为关键帧,生成视频摘要.实验结果表明,文中算法生成的视频摘要内容覆盖率高,可以根据视频长度和类型灵活直观的确定关键帧数目,并且对于不同类型和长度的视频给出了累计贡献率的取值范围,可以为用户提取合适长度的视频摘要提供有效依据.
[Abstract]:In order to facilitate the browsing and storage of video, video summary technology for the purpose of summarizing video content has become very important. Aiming at the problem of flexibility and indirectness of the two methods of determining the number of key frames according to prior knowledge and threshold adjustment in video summarization at present, A video summarization algorithm based on SVD and sparse subspace clustering is proposed. The algorithm uses SVD to reduce the dimension of video in time dimension, determines the number of principal components of video in time dimension, and takes the number of principal components as the number of key frames according to the cumulative contribution rate. Then the sparse subspace clustering algorithm is used to cluster the video frames. Finally, the frames with the largest correlation with other video frames are selected as key frames in each category to generate the video summary. The experimental results show that the proposed algorithm has high content coverage and can determine the number of key frames flexibly and intuitively according to the length and type of video, and gives the range of cumulative contribution rate for different types and lengths of video. It can provide an effective basis for users to extract video abstracts of appropriate length.
【作者单位】：西北工业大学理学院;
【基金】：国家自然科学基金(61070233)
【分类号】：TP391.41

【相似文献】