Video Copy Detection Based On The Expression Of Spatiotemporal Information

With the development of the Internet and multimedia technology, the amount of online videos has grown tremendously in recent years. Massive video applications bring the difficulties in mass memory, video retrieval and video copyright protection. Content-based video copy detection (CBCD) is proposed to solve these problems. This paper introduces the framework of video copy detection and some kinds of video copy detection technologies. And we summarize four key technologies:shot detection, key frame extraction, feature extraction and feature matching. Inspired by prior works, we propose two new algorithms for CBCD.1. A video copy detection method based on median of key frame is proposed. This method extracts the key frames from a video, then divides each key frame into several blocks and extracts the features based on the median of each block. As the median is less sensitive to global transform than the average intensities, this method has better performance than those using the features based on the average intensities. To localize the copy location in the query video and the reference video, we propose a variable-length sliding window method for matching and localization, which can localize the beginning location in the query video and the reference video. And the length of the truth copy video can also be evaluated accurately. The experimental results show the proposed approach has a good performance in terms of detection accuracy ratios and also can provide precisely temporal localization.2. An effective and efficient CBCD method is proposed, which is based on spatial feature and the transform of the spatial feature with the time-variation. Firstly, the key frames are extracted from a video. Secondly, we calculate spatial gradient feature and temporal gradient feature for each frame, and the spatiotemporal gradient feature can be extracted by fusing the spatial gradient feature and the temporal gradient feature. Finally, a similar list for each query frame is constructed based on the similarity of the proposed feature and the copy detection result can be acquired by fusing the similar list under the condition of temporal consistency. The experimental results show the proposed scheme is effective in both copy detection and beginning localization.