Context-Constrained Near-Duplicate Video Retrieval

With the development of the Internet, the number of web video has increased tremendously. Analysis on leading social video sharing websites reveals a high amount of videos with redundant content. Users are often confused when they have to go through a large number of redundant videos before they find what they really want. Near-duplicate web video detection can be used to eliminate the redundancy to improve the user experience. Near-duplicate video detection can also be used to other applications, such as automatic video annotation, subject tracking, copyright protection.Existing near-duplicate video detection methods either pursue the retrieval accuracy excessively while ignoring the efficiency, or only pursue the retrieval efficiency while sacrificing the ability of near-duplicate video detection with complex change. This paper proposes a method to combine the global and local features with the constraint of contextual information, which has a good trade-off between retrieval accuracy and efficiency. First, we use a gray level based global feature to find some near-duplicate video with a high degree of confidence. Second, we use the time duration and the number of views to find the seed videos. Third, we use the title, tags, time duration to filter most videos to get a candidate video set. Finally, we use the local features of seed video to detect the candidate near-duplicate video set through more costly computation.Experiments on an open web video data set demonstrate that the proposed method outperforms the gray level based method on accuracy. Compared to the method with the combination of gray level based global feature and SIFT based local feature, our proposed approach can achieve228times in speed with just a slight loss of accuracy.