Research On Shot Detection And Scene Detection In Content-Based Video Retrieval

With the development of computer, network and multimedia technologies, digital video has been used more and more widely. How to retrieve interesting data rapidly and effectively from the extremely large video data has become the key problem. Content-based video retrieval has been developed and become the hot technique of multimedia research in recent years. In this paper, Shot detection and scene detectoin in content-based video retrieval are studied technically, and some results are obtained as follows:Shot detection algorithms are analyzed systematically, and the color feature and texture feature of frames are extracted based on wavelet transform, the mutual information of color feature and the co-occurrence mutual information of texture feature are colligated to detect cut shot boundary.A wipe detection algorithm is presented in this paper. In wipe, the frames of the first shot are covered gradually by the frames of the second shot, according to this feature, candidate wipe can be detected. Line or ellipse can be detected during the transition by Hough transition, and some features of the line or ellipse change regularly, based on which, the wipe can be conformed.Dissolve and wipe detection algorithms based on editing model are proposed in this paper. During dissolve transition, the first difference and second order difference of the variance change regularly, and the candidate dissolve regions are detected based on this character, but some motion of object and video camera can cause such curve, so the candidate dissolve regions are conformed based on the character of successive monotonic change.In this paper, wipe is divided into linear wipe and secondary wipe, and they are detected respectively, and the results show the high performance and efficiency of the algorithm.Dynamic frame is defined in this paper to extract key frames from shots, and the construction methods of each gray value of the pixel in dynamic frame if given. The key frames are determined by the distance between the frame and the dynamic frame. Results show the algorithm has high fidelity and compression ratio.It is more difficult to detect scene than shot boundary, which is because shot is a phisical partition of video, but scene is a semantic partition of video. A scene detection algorithm is present in this paper, which is based on the extraction of the key frames. In the algorithm, shots are clustered firstly, and the correlation coefficient between the classes of shots is calculated, and then, the classes of shot is combined to form scene, at last, the scene whose shots is less than three is combined with its adjacent scene. Results show the scenes detected by the algorithm is close to that detected by artificial judgement.