Research On Video Key Frame Selection And Visual Representation

With the development of computer graphics and Internet technology, large amount of video and animation has been created and spread through Internet. Key frames are a sequence of discrete frames which contain significant visual content of original video. Key-frames-based approaches serve as important methods utilized in video retrieval and high efficiency animation rendering. Key frame selection strategy, which is valuable on both theoretical research and application, turns into a crucial research task in recent years.The paper firstly proposes the definition of content distance between two frames. We exploit pixels color values and their distributions as the basic feature of frame and use them to calculate the content distance. Specifically, we make use of the moments of pixels color values in each frame and the correlate coefficient between two frames to define one content distance. Moreover, according to information entropy theory, we exploit joint entropy and mutual information, and relative entropy between two frames to give two other content distances.We depict a key frame selection algorithm motivated by maximize the video content information presented by key frames. The algorithm is non-parameter which allows users, rather than the syntax structure of the original video, decide the quantity of key frames. The algorithm also provides a unified process for key frame selection and video partition. We test the algorithm with different video and different amount of key frames. Experiment results demonstrate that our algorithm performs better than two other key frames selection algorithm.We propose a time order based key frames visual representation, through which users can understand the segment structure clearly and observe the key frame in each segment interactively. Further more, we propose two kinds of non-time order based visual representation of key frames: exploratory method and guided method. The successive key frames in exploratory method have distinctive visual content while in guided method they share similar content. We also give a novel visualization model to show these two methods vividly and interactively.