Research And Implementation Of A News Video Abstract System Based On Content

With the rapid development of multimedia technology and network technology, video media, as one of the major information carriers, has tightly associated with human working and living. The characteristics of video, such as huge amount of video files and long duration, make it hard for highly effective organization and management. Because of video’s unstructured data formats and hardly described content, we have difficult to quickly browse and retrieve a certain part of content. Content-based video retrieval (CBR) is efficient way to solve this problem. Since news video is an important type of video media, which has high value of study, the thesis focuses on news video media.In the thesis we study the key technologies of news video abstract, including shots segmentation, key frames extraction, story unit division, mouth broadcast frame detection. For shots segmentation, we calculate the histogram of every frame of video based on HSV color model and compare the difference between neighboring frames to tell whether the switch occurs. Fixed threshold is a big problem because each video has its own threshold and each shot has various thresholds when the content changes greatly. So we propose an adaptive threshold method to solve this problem. For key frames extraction, we put forward an adaptive algorithm with various numbers of frames in order to make the key frames more comprehensive and more representative. For story units division, mouth broadcast frames detection is used to get the result since it is one of the unique structural features of news video. In the mouth broadcast frame detection process, we combine the face detection method and data clustering method to achieve better result.At last, a prototype system is implemented to analyze the structure of news videos and get the abstract of them. Users can use it to browse long duration news videos in a short time.