News video story segmentation

Abstract

This paper presents a two-level framework for news video segmentation. Our framework is established-based upon a similar framework as in. We extended the original framework by adding rule-based pre-segmentation module, similarity measurement module and new features. We perform decision tree at the shot level and HMM (hidden Markov models) analysis at the story level, respectively. Experiment result with a training set of 24 hours (967 story units) news video from CCTV-9 (China Central TV-International) and a testing set of 24 hours (779 story units) news video from several TV-channels show that our semi-automatic system can achieve 81.5% of F1 value in the case of CCTV-9