Research On Key Techniques Of Content-based Video Retrieval

With the development of computer, communication and multimedia technologies, the amount of video data is enlarged at an explosive speed. How to organize, manage, express and retrieve these video data efficiently becomes a hot issue. This paper imports the theory of information into video retrieval, and conducts thorough research in shot boundary detection and key frame extraction. Firstly, based on the systemic summary of the available shot boundary detection algorithm, a general shot boundary detection algorithm was proposed. This algorithm combines mutual information with histogram as the similarity of two image frames, and then it adopts a dynamic window to unify abrupt transition and gradual transition for universal processing. Experimental results show this algorithm can improve the recall rate and precision rate at the same time. Secondly, based on the combination of visual matching and clustering approach, a new mutual information based key frame extraction algorithm was introduced, which can extract one or more key frames from one shot by the specific character. Experimental results show this approach has good performance on key frames according to those by manual work. Finally, a tentative work about key frame extraction based unsupervised clustering method was proposed. This approach can extract key frames from video sequence directly and does not require any pretreatment such as shot segmentation. Computer simulations show this algorithm can give a good representation of video content and make video retrieval much easier.