Video Key Frame Extraction through Canonical Correlation Analysis and Graph Modularity

Key frame based video summarization has emerged as an important area of multimedia research in recent times. In this paper, we propose a novel automated approach for video key frame extraction in compressed domain using canonical correlation analysis (CCA) and graph modularity. We prune certain edges from the Video Similarity Graph (VSG) using an iterative strategy until there is no improvement in graph modularity. Resulting connected components in the final VSG correspond to separate clusters. The proposed algorithm also uses multi-feature fusion using canonical correlation analysis to achieve higher semantic dependency between different video frames. Experimental results on some standard videos of different genre clearly indicate the superiority of the proposed method in terms of the F 1 measure.