Study And Application On News Video Segmentation Based On Visual-audio Features

With the rapid development of multimedia and network technique, effective analysis and utilization of the video data become more and more important. Content-Based Video Retrieval(CBVR) then is becoming one of the research focuses. News video, as a kind of video data, owning to the intensive information and extensive coverage, now is one of the important information sources, and Content-Based News video retrieval is an essential part of the CBVR.Compared to other kinds of video data, news video has many major difference: clear semantic hierarchy and semantic identification, fixed anchorpersons, simple switching mode and so on, all these characteristics make the segmentation of news video differs from other kind of video data.Firstly, video text detection and segmentation are studied , and a gradient reinforcement-based adaptive news video text segmentation is presented. The experimental result proves a better text recognization rate can be get in this method.Secondly, the necessity of the audio features in the news video segmentation is analyzed, a short-term energy-based method is implemented to detect the silence interval, a Mel-cepstrum and Gaussian Model-based method is presented and implemented to detect the speaker changes. Audio and visual features are combined to segment the news video, and a better accuracy rate is obtainted in the experiments.Finally, a prototype system based on visual-audio features is implemented. Visual and audio features such as headline, anchorperson shot, silence interval and speaker change are used in the system to dectect the partition points of new envents. CCTV news video can be automatically divided into certain number of new events in this system, at the same time the beginning , ending time and some semantic information such as caption texts and key frame are extracted and stored into the database, thus, keyword-based news information retrieval and browsing can be provided in this system.