News Video Clip Retrieval Based on Topic Caption Text and Audio Information

This paper presents a novel scheme for news video structure indexing and news story clip retrieving. Instead of using low-level features, the method is built upon the combination of topic content and visual features. First of all, a new method of topic caption text detection is proposed in which the topic caption frame is detected by features extraction from frame differences, the topic caption lasting time and times of caption transition in the same shot, and the dynamic split-merge strategy is used to segment individual character. Afterwards, one news video is segmented into a series of news story clips on the basis of topic caption text and silence clip. Finally, a reasonable model is built to retrieve news story clips by both visual similar degree and topic content relativity between two news clips with strong semantic meaning. The experimental results showed that the proposed method could effectively detect and recognize topic caption text and index news story clips.