Video Copy Detection Based On Multiple Visual Feature Synthesizing Method

With the rapid development of computer, network and multimedia technology, the number of digital video multiplied. The videos on Internet are often edited, copied, and reproduced without any supervision, and become lots of copy version. This has brought great challenge to copyright protection. How to detect video copy automatically in mess data has become a technical problem to be solved urgently. Video copy detection is the essential problem in many application, such as web search improvement, advertisement monitoring, video retrieval based on instance and concept trace. Recently, researchers has been doing lots of work on content-based video copy detection, however the performances are less than satisfaction.One of the essential problem in video copy detection is how to improve the robustness to various transformation, and the research should been focused on finding a proper feature which can describe the invariance of video content; the other essential problem in video copy detection is how to improve the precision of the detection, and the research should been focused on finding a proper similarity matching algorithm.To solve the problem we discussed above, we proposed a novel method which employed multiple visual feature synthesizing to improve the system robustness to diverse transformation, and used variety of optimization methods to improve the accuracy of retrieval. The experiment results show that this solutions has very strong robustness for a variety of copy transformation, and can get a better precision. Based on these, This paper further put forward to discriminate the type of copy transformation, and used this information to further elaborate the results of retrieval. Good experimental results have been achieved.The main work of this paper is as follows:1. Due to the uncertainty and diversity of the type of copy transformation, it is difficult to achieve great performance based on one single visual feature in video copy detection. This paper proposed a novel method which employed multiple visual feature synthesizing method to solve the problem. We employed traditional local and global visual features, and this strategy improved the system robustness to different copy transformation, and achieved better recall rate, and it mainly solves the key frames coarse retrieval in massive video data efficiently.2. This paper also put forward the accurate similarity measure of video copy, including two aspects:key frame content similarity measure and video similarity measure. In key frame content similarity measure, we used NBS-based NCC to measure the key frame content more subtly. It filtered out invalid key frame results which the coarse search returned, in order to improve the result of copy video key frame content similarity matching. In video similarity measure, we employed a improved frame sequence matching method based on time sequence consistency to get final video level result. This approach can better distinguish between correct results and errors, in order to reduce the rate of misstatement, and to improve the detection accuracy.3. Finally, the paper also studied the discriminant of copy transformation.We should not only retrieve the video copy precisely in huge amounts of dataset, but also to discriminate the transformation which generated the copies of video. This can lay a foundation for pirate evidence obtaining in many applications such as intellectual property protection, and prepare for mining further relationship between different copy transformation and feature selection. Finally on this basis, we can further optimize the results for video copy detection.