Research On Fast Video Shot Boundary Detection

Video is an important visual information media and it has richer and more intuitive content than text, audio and image. With the development of computer science and Internet technology, video information has been increasing rapidly. The vast amounts of video information make high requirements on the video information content management technology. Video shot detection is the first and essential step for video content management technology, and it has great influences on the performances of content-based video retrieval, summarization, content identification and other applications. Thus it has strong research significance. In this thesis, we consider the real-time algorithm as important concern and research on fast shot boundary algorithms. The aim of the thesis is to design an efficient algorithm to improve detection speed and accuracy, try to solve real-time application requirements.In this thesis the existing methods are fully studied, the obstacle of detection speed is analyzed firstly. Then we focus on candidate segment selection and singular value decomposition dimension reduction in order to improve detection speed. And then the fast shot boundary detection algorithm is introduced. For cut transition detection, we design adaptive thresholds and detect feature differences of two consecutive frames. For gradual transition detection, a unified gradual transition detector is proposed. The algorithm constructs a new metric and detects gradual transitions by graph pattern matching. Meanwhile, the algorithm considers the optimization of detect results and necessity cross detection. Finally, we do experiments on TRECVID2001test data and other video materials. By comprising with other algorithms, the results show that the proposed scheme can accelerate detection significantly and achieve excellent accuracy.