Multi-scale information maximization based visual attention modeling for video summarization

The efficient browsing and retrieval of videos is a fundamental part of current multimedia systems especially on mobile devices. One way to provide enjoyable video browsing to the users is by providing summaries of the videos whereby the users can have a clue of the contents of the video before watching the video itself. In this paper, we propose a key frame extraction based video summarization technique based on user attention modeling. Our technique builds static and dynamic attention models of human perception to predict those frames which are visually salient for the humans. For the static attention clue, we propose the multi-scale version of the information theory based attention modeling. For dynamic attention modeling, the motion intensity has been used. The experimental results and comparison with other techniques shows that the proposed technique yields efficient results.