一种基于K-L变换和聚类的视频摘要方法

Video abstraction based on K-L transform and clustering

提出了一种基于K-L变换和聚类的视频摘要方法。首先通过对视频帧原始RGB空间进行K-L变换,得到由主轴构成的参数模型;其次运用滑动窗口法进行镜头检测;再次,根据最邻近规则对每个镜头的视频帧进行聚类;最后通过后处理优化聚类结果,提取最靠近聚类中心的帧作为关键帧,组成视频摘要。以新闻视频为例,实验结果证明了算法的有效性。

This paper proposed a new approach of videoion based on K-L transform and clustering. Firstly,the algorithm used K-L transform of the RGB color space to produce parameter model that was consisted of eigenvectors. Secondly,the algorithm used a fixed length sliding window of current frame to detect shot boundary. Thirdly used the nearest neighboring rule to cluster the frames in each shot. Finally,used the post treatment to optimize the results,and frames which were most nearest to the center of the clusters are extracted as key frames,which constituted video abstraction. Abstraction is generated for news video,and the experimental results prove its efficiency.