Video Summarization via Simultaneous Block Sparse Representation

Abstract

With the ever increasing volume of video content, efficient and effective video summarization (VS) techniques are urgently demanded to manage the large amount of video data. Recent developments on sparse representation based approaches have demonstrated promising results for VS. While most existing approaches treat each frame independently, in this paper, the block-sparsity, which means the keyframes or non-keyframes occur in blocks due to the content similarity in a same frame block, is taken into account. Therefore, the video summarization problem is formulated as a simultaneous block sparse representation model. For model optimization, simultaneous block orthogonal matching pursuit (SBOMP) algorithms are designed to extract keyframes. Experimental results on a benchmark dataset with various types of videos demonstrate that the proposed algorithms can not only outperform the state of the art, but also reduce the probability of selecting non-informative frames and "outlier"frames.