Research on News Video Story Correlation Analysis
|School||National University of Defense Science and Technology|
|Course||Control Science and Engineering|
|Keywords||News video Story correlation analysis Near-duplicate keyframe matching Bag-of-words model Co-clustering|
With the rapid development of IT and network technology, news communication industry is developing by leaps and bounds, and video news becomes the most prevalent Internet value-added service applications. In the face of massive amounts of news videos, users are usually concerned about of the information relating to a particular event. So the method of effectively identifying the similar videos is sorely needed. While news videos have complex structure and diverse content, it’s a challenging work to find the news videos concerning same event in mass video data. In this paper, we explore and study this subject which has important theoretical significance and broad application prospect. We are aiming at solving some key technologies on the news story correlation analysis, so as to provide technology supports for semantic content analysis of news videos.Through analyzing the relevant works, this paper first presents a processing flow of news video story correlation analysis. The realizing technical approaches and some key technologies involved are discussed, in an attempt to make clear the main task and problems of our research work. We study the preprocessing technologies of news video story correlation analysis, and improve the algorithm of story segmentation and keyframe extraction to a certain extent.This paper puts forward a method of near-duplicate keyframes matching based on the bag-of-words model. The involved technology, such as the determining of the clustering quantity of key points, reducing of the dimension of visual vocabulary and assigning of the weight of visual words have been expounded in the paper.An approach of two-stage story correlation analysis technology based on text and visual character fusion is presented. In first stage, we cluster the joint matrix utilizing co-clustering methods after an integration of text-story probability matrix and visual-story probability matrix in order to get the clustering groups of similar stories. In second stage, optimization of the outcome of first stage based on weighted sum of textural similarity and visual similarity between the stories is performed.We verified our methods by experiments and analyze the result. On these bases, a prototype system of news video correlation analyzing is designed and implemented. The system integrates the fruits of this paper, it offers an events-oriented analyzing and organizing scheme for news video database, and provides elementary technologies to support efficient browsing and retrieval on the news video data.