Classification Program and Story Boundaries Segmentation in TV News Broadcast Videos via Deep Convolutional Neural Network

Given the amount of video information on the net, the user has had difficulty finding the information in a reasonable amount of time. Thus, all video content must be segmented and annotated so that he/she can access the information directly. The goal of the proposed approach is to allow a better exploitation of video by multimedia services (TV-On- Demand, catch-up TV), social community and video-sharing platforms (Youtube, Facebook…). In this work, an approach to classify TV programs and story boundaries segmentation in TV news broadcast video using Deep Convolutional Neural Network (DCNN) is presented. The first step is to extract features from video. This characteristics will modeled as video corpus governings the organization of TV stream content. This organization is carried out on two levels. The first consists in the identification of anchorperson by Single-Linkage Clustering through CNN faces and the second level aims to identify the story of news program due to the large audience because of the pertinent information they contain. In addition, we implement a 360-h broadcast video dataset obtained from five French news channels with ground-truth marked semantic shot categories, program genres and story boundaries. Experiments on this dataset prove the relevance of our approach for news broadcast video segmentation.