Hierachical Structure Based Video Summary

As the development of multimedia information technology, various productions of this kind of techniques have emerged, especially, a variety of digital video. But how to find a video according our requirements and satisfying the timeliness and accuracy in the ocean of video is a challenging problem. Therefore, the video summary has appeared to solve the problem. Video summary is first to analyze the structure and content of video in a automatic or semi-automatic mode, then extract meaningful frames from original video exploiting the analysis result, and finally organize these frames in a certain way in order to form a summary which covers the general information of the analyzed video. The video type can be news videos, entertainment videos, TV shows, movies and other forms, and the video summary can also have a plenty of forms. These include mostly used static video summary (a set of the key frames), dynamic thumbnail video (contain sound, music, images and other short video), many different forms of video posters and so on.This thesis mainly focuses on static video summary. In the course of the study, concentrated on different types of video (such as common video, news and talk shows video), based on low-level features, from the angles of the video content, interaction strength and intensity change we analyze and generate video summary which covers the rich content and expresses certain semantics of the video. The key technique under study is the video feature extraction which is suitable for video summary.According to the ordinary video summary, the research starts from the bottom video features, segment the shot in the HSV space efficiently, use the dynamic sliding window method to remove redundant frame to generate an effective summary with less redundant and human interaction is not needed in the process.For news video, mainly from the perspectives of host detection, title and subtitle detection, the color matching of host’s clothes, and its unique hierarchical structure from image frames, shots, story units to the whole news. The shot as the analysis unit can reduce complexity of the analysis and is the accurate and complete partition of each story unit. The result can better express the news video content.For talk show video, by studying the temporal characteristics of the main colors, the semantic of interaction strength in the chat show is proposed. When analyze the similarity inside the shot, intensity change probability during a period of time is proposed using mutual information changes in the dynamic regional. Combining the interaction strength and dynamic regional similarity, the generated summary is capable of expressing these two kinds of semantics. In order to adaptively extract sampling frame within a shot, the extraction strategy adopted is the least similarity of the global changes in the shots. The video summary is able to satisfy our needs.At the end of this thesis all the work and efforts about video summary technology are analyzed and summarized, and the possible improvements for the existing problems are described.