Research On Content-Based News Video Abstraction Technology

With the rapid development of communication technology and Wide Band technology, the amount of video data increases unprecedentedly. However, the efficiency of video browsing is very low due to its nonlinear and unstructured data form. How to automatically establish index structure which can analyze and manage the video streaming and locate those user interested contents in the giant video data? Researchers pay close attention to these problems. News video belongs to a kind of special video, which can provide important information of current events. Content-based news video abstraction technology can reduce the video data volume, improve the searching efficiency, save the browsing time and have important significance on the various applications of news video.This paper mainly focuses on the content-based news video abstraction technology, and the main achievements are given as following:(1) In static video abstraction, a new method based on K-L transform and clustering is proposed. Firstly, K-L transform of the RGB color space is used to produce parameter model that is consisted of eigenvectors. Secondly, shot boundary is detected with a fixed length sliding window of current frame. Thirdly, frames in each shot are clustered by the nearest neighboring rule. Finally, post treatment is used to optimize the results, and frames that are most nearest to the center of the clusters are extracted as key frames to constitute video abstraction. The experimental results show that this method is efficient with favorable commonality and flexibility.The keyframe sequence extracted can better reflect the vedio content.(2) In anchorperson shot dectection, a new method based on rules is proposed. Firstly, extended facial regions (EFR) are extracted and clustered for clothing of anchorperson in a single news video is not changed usually. Secondly, three judgment standards are given according to anchorperson's rules of occurrence time. Finally, anchorperson shots are confirmed with massive narrowing down process. The experimental results show that the new method is efficient with high popularity and detecting accuracy.(3) In dynamic video abstraction, a client-oriented method for news video skims is proposed. The method provides three compression strategies for users, including compression based on anchorperson shot detection, fixed ratio compression and compression based on caption detection. In each strategy, users'need is considered to some extent, and video skim is generated with human-computer interaction. The experimental results manifest that the news video skim can satisfy different customers'requirements well.