News Video Story Segmentation Based on Naïve Bayes Model

Abstract

Story boundary detection is the foundation of content based news video retrieval. In this paper, Naive Bayes Model, which has been successfully used in multi-modal feature fusion, is implemented in news video story segmentation. Firstly, we get candidate boundaries through shot detection. Secondly, middle-level features such as visual features, audio type, motion and caption, are extracted from shots around these boundaries to generate input attribute set of the model. Thirdly, we use trained Naive Bayes Model to compute posterior probabilities that a candidate boundary is a real story or not, and get the result according to maximum posterior probability rule. Lastly, post-processing is conducted, removing the non-news stories. Experiment results show that this method is effective and achieves satisfactory precision and recall. The new method requires less computation and is applicable to different types of news programs.