Video Copy Detection Research Based On ORB Feature

Network videos’rapid development and popularization brings forth the problems of the digital video copyright. It is video copy detection technology that its task is to detect whether there is a copy in the video clips, so video copy detection technology research plays a vital role for video copyright protection and management.There are two main methods of video copy detection technology, one is based on content and the other is based on digital watermarking technology. The method based on content is better than the method based on digital watermarking technology. The reason is because that the method based on content without any embedded information into the video, it can directly extract robust features from video content itself, so it is convenient and the detection result shows better. Therefore, we choose video copy detection method based on content to conduct research. The basic idea of this method is achieving video copy detection through extracting feature from video key frame and feature matching. There are mainly three steps:video preprocessing, feature extraction and feature matching. The key problem is to find video features of robust to various copy attack. Many researchers have given more focus on the classic floating-point local characteristics such as SIFT descriptor, SURF descriptor and global feature such as ordinal measure feature, GIST and so on. A few excellent binary local characteristics such as BRIEF, ORB and so on are put forward in recent years. These features can detect a variety of copy attack, and the memory consumption, the matching accuracy and speed of matching show better performance the floating-point local characteristics. In this paper, we propose taking ORB (Oriented FAST and Rotated BRIEF) binary feature descriptor local characteristics into account and use it to study video copy detection.The other main contribution of this paper is to put forward a video frame feature. It is a rapid global feature named Hilbert features based on key points, which combines Hilbert curve with local key points of ORB characteristics. Although the characteristics of robustness is not very strong, it owns fast detection speed and can guarantee the advantage of higher recall rate. On this basis, this paper combines the local characteristics of robustness advantages with global features of the detection speed advantages, then propose a kind of based on global features and local features fusion video copy detection method which is the ORB binary local feature descriptor, global ordinal measure feature and Hilbert feature based on the key points combine together. To begin with in video preprocessing, computing adjacent video frame Bhattacharyya distance of gray histogram for video shot segmentation, and choose the first frame as a video shot key frame, and further remove repeating frames to get new key frame sequence. What’s more, extracting ordinal measure feature, Hilbert feature based on the key points and ORB feature from video key frames respectively. In the matching step, using ordinal measure feature with the sliding window method to query video for the first match which can remove part of the unrelated video. And then using Hilbert feature based on the key points with public video frames match again by gray sequence detection results. Through two fast filtering by the global feature, we maintain a high recall rate. Finally, use the ORB feature accurately matching to get the final test results from the two filter results. In order to improve the speed of detection, we propose a method of index for video features which is similar to local sensitive hash.Experiments show that comparing with other methods, the proposed method keep a high recall rate and have good precision at the same time whether using index or without index. Though the retrieval time gets a promotion without index, it greatly improves when using index.