Research On Video Shot Boundary Detection And Key-frame Extraction

The advances in computer, network and digital video technology have resulted in an explosion of digital video data in the last few years. The organization, summarization and retrieval of video data have become the focus research.In this thesis, we focus on the two basic technologies of contend-based video retrieval, which are shot boundary detection and key-frame extraction. Firstly, a new algorithm based on the co-histogram between two consecutive frames is presented. In the algorithm, the similarity is defined based on the co-histogram symmetry which is high in the same shot, but low in a shot transition. The similarity effectively reveals the transition characteristics. The difference of similarities is used to enhance the abrupt transition features, and an effective bidirectional searching flashlight detection approach to distinguish flashlight scenes from abrupt transitions. A Finite-State Automata is used to improve the robustness in detecting gradual transitions. Experiment results show a good performance in recall and precision. Secondly, a novel approach based on co-histogram to extract key-frame is proposed. The approach uses the symmetry of co-histogram between consecutive frames to define the dissimilarity, and classifies frames by detecting the high slope points in a cumulate curve of the dissimilarity. Finally, Key frames are dynamically extracted from key classes. The experiment results show that the approach achieves a high compression rate and an efficient summarization.