Web Content Security Analysis And Algorithm

This thesis presents a framework for news video story segmentation. The framework consist of three layers: feature extraction layer, shot tagging layer and story boundary detection layer. At first layer, a tandem feature extraction method is implemented including: audio classification, face detection, caption detection, scene change detection, and speaker change detection, and so on. And then at second layer, with these high-level features together with other low-level features, the Decision Tree is employed to classify the shots into predefined categorizes. Finally at third layer, HMM-based technique is used to perform maximum likelihood estimation of the story boundary. To more efficiently represent the patterns of the stories, some new features including "Topic Feature", "Face Number" and "Face Position" are added into the framework, and also a similarity measure technique and pre-segmentation technique are used to improve the performance of framework. Experiment result with a testing set of news video clip from different channel show that the semi-automatic system, which is based on the framework, can achieve 71.9% of average Fl value, and 81.5% Fl in the case of CCTV-9.