Research And Implement Of Video Retrieval System Based On SVM

With the rapid development of video decode & encode and Internet, the traditional video retrieval system cannot match the people's need, so the CBVR(Content-based Video Retrieval) became the hotspot in video research, mainly focused on video content retrieval , video content match, video feature match and so on.This paper propose a new video retrieval system framework and explain our research in key algorithm of video retrieval. Main works of this paper are summarized as follows.1. Traditional frameworks for video retrieval are mainly focused on key algorithms, but not study the procedure of video retrieval system, this paper mentioned a new video retrieval system framework after combined SVM algorithm and video retrieval, and the new framework can works effective in a simple way.2. This paper summarized the key frame selection methods, and compare them with the temporal sampling method. Experiment result showed us the temporal sampling method is better than the other methods. Then we optimize the temporal sampling algorithm with Dijkstra algorithm, reduce the time complexity from O(n~4) to O(n~2)O n 2, and space complexity from to O(n).3. This paper use the MPEG-7 visual descriptor to extract the video feature, and study the impact of using specific feature classifier in semantic conversion. Experiment result showed the positive effect of using specific feature classifier(like face detector).4. After some study on relevance feedback, boosting and result re-ranking, we propose a new SVM semantic conversion procedure. The new SVM semantic conversion procedure can have a increase of 16.2% in average precision and 12% in recall factor.5. A video retrieval system was developed based on this paper's theory, we can easy to exchange the video retrieval system data based on this platform. So this system has some practice meanings.Finally, the conclusion of this paper was made and some future works are presented.