Video summarization by video structure analysis and graph optimization

We propose a novel video summarization method that combines video structure analysis and graph optimization. First, we analyze the structure of the video, find the boundaries of video scenes, then we calculate each scene´s skimming length based on its structure and content entropy. Second, we define a spatial-temporal dissimilarity function between video shots and model each video scene as a graph, then find each scene´s optimal skimming in the graph with dynamic programming. Finally, the whole video´s skimming is obtained by concatenating the skimmings of the scenes. Experimental results show that our approach preserves the scene level structure and ensures balanced coverage of the major contents of the original video.