With the development of multimedia and the Internet, videos increase rapidly. Because of a mass of videos, it is necessary to automatically process, analyze, and retrieve information by means of computer technology. Content-based video retrieval has become hot off the press, in which shot detection and copy detection are two important parts.Shot detection is the prime task in video retrieval, which directly affects the results of retrieval. It is important to achieve the inherent features of video streams and estimate their classes. This paper first introduces the label propagation algorithm used in classification problems. Each frame is regarded as a sample nod of label propagation. Then, the inherent features of the shots are obtained using the semi-supervised learning. Our proposed algorithm is different with the existing methods, and better reveals the essences of video data.For copy detection, this paper combines several kinds of features to extract the key points feature vectors of the clips. In order to resist the changes induced by shooting angle, shooting time and variant illumination, SIFT(Scale Invariant Feature Transform) descriptor is introduced as a local feature. The shape context algorithm is combined with the corner points to emphasize the video contents. And the application of local corner points simplifies the calculations.In the model of copy detection, multiple measure factors are applied to determine the similarity. And a weighted common subsequence is designed to judge the ranking factor, which is more feasible than the traditional longest common subsequence. |