TV News Retrieval Based on Story Segmentation and Concept Association

Abstract

In this paper we propose a novel method for TV news retrieval. A first stage concerns a temporal segmentation into stories units. Then, for each story the most relevant concepts are extracted based on a multimodal fusion between visual and textual information. By analyzing the video stream, we perform global frame representation, image retrieval and re-ranking, in order to determine, with high confidence, the segments boundaries. In addition, by using the video subtitle, we identify the most relevant concepts / topics addressed in each independent segment. The framework is evaluated using one week video archive of France Television and 20 journals from NBC and CNN TV stations. For the temporal video segmentation, our system returns high precision and recall scores, superior to 90%. Regarding the topic association technique, we obtain a mean average precision score superior to 0.5.