Automatic TV Program Segmentation And Similar Video Retrieval

With the popular application of network TV and Digtal TV, automatic TV program segmentation from TV stream and similar video retrieval are very important research topics for users'varios demands, TV content supervising and video retrieval. The work and innovation of the dissertation can be listed as following:By analyzing the characteristic of TV program structure, TV commercials detection algorithm based on audio and visual features according mute frames, shot boundaries and length of commericals is proposed. Experimental results show this method can meet accurate rate higher than 90%. Accoring to the properties of TV program repetition, the repeat detection approach based on multigranularity histogram comparison with video boundaries adapatation is proposed. To improve computational efficiency and noise resistance, video copy detection method based on video signature with ordinal measure was also proposed. TV program segmentation algorithm based on repeat detection combined with commercial detection according to the repetition of TV program is also proposed. Experimental results show these algorithms are effective with accrate rate higher than 95% for repeat detection.To improve the computation efficiency and accurate video similarity measure, video feature extraction approaches based on video signature and video component proposed according to the statistics of spatial-temparal distribution. The video similarity is measured by the distance of video signature and the number of similar video components. Video feature extraction can be elemented efficiently and video similarity can be effectively measured with accurate rate higher than 94% with the computational efficiency faster than 40% with the compared algorithms.To achieve fast video retrieval in large video database, Clustering index table baed on video signature and video component is proposed combining video feature and index according to the character of video feature extraction and video similarity measure. Experimental results indicate these algorithms can meet the requirement of scalable computing with the satisfying recall and precision rate higher than 95% and search time within 0.01s for ten thousand video clips under common PC environment.With the algorithms of video feature extraction, similarity measure and fast search based on clustering index table, a video similarity search system was implemented. Users can access the video by key words search and content-based video similarity search. Video-on-Demand service is also provided by video streaming technology. Video similarity search system has very high search efficiency and can meet users'different needs of video similarity search and it is also applied into uask video search system supported by 863 Program.