An unsupervised approach for story-related subject caption detection in broadcast news video

The story-related subject caption (SSC) in broadcast news video expresses the subject of news story, and plays an important role in news story segmentation and news video indexing. We find that a SSC always has a strip background and all the SSCs in one news video have the same style. By taking advantage of these characters, this paper presents an unsupervised approach to detect SSCs in broadcast news video. We first filter out most of the frames without SSCs by detecting horizontal lines. Secondly, classic text detection technique is utilized to detect captions on frames with horizontal lines. At the same time, spatiotemporal slices processing is employed to track the detected captions and avoid rescanning. Thirdly, all the detected captions from above steps are treated as candidate captions, and clustered by spectral clustering. Finally, according to the caption clusters´ amounts and spanning times, we select one cluster of captions as SSCs. Experimental results show that the proposed approach can detect SSCs in broadcast news video accurately.