The Research Of Video Copy Detection Based On Frame Fusion Framework

With the exponential growth of online videos on the internet, CBCD is becoming more and more important for digital copyright protection, media content tracking, cross media retrieval, and so on. CBCD aims at deciding whether a transformed query video or part of this video is a copy of a video in reference dataset and where the copy is originated from. Two popular video copy detection schemes include sequence matching based and frame fusion based frameworks. Since the sequence matching based method treats a fixed-window as detection unit, it cannot smartly handle the detection problem of multiple video copy clips in unbounded video stream. In contrast, the frame fusion based scheme can address this problem well by searching the reference database and fusing the returned lists of similar reference frames. However, the computational complexity of frame fusion based method is high since a large amount of returned reference frames need to be handled. In addition, a classical uniform frame sampling strategies generally used for improving copy localization precision not only further aggravates the computational efficiency but also leads to much more false alarms due to content redundancy among frames. Moreover, the outputs of frame fusion system are some paths which reflect the location of copied video, yet not all the paths are true ones. Therefore, some path removing algorithm is required to filter out false paths. The main effort of this thesis is to alleviate the above problems. The main contributions include:1. We propose two frame filtering strategy, which are based on the similarity of adjacent frames. The first strategy describes each image with VLAD descriptor, and then dynamically selects frame pairs needed to be compared. The second strategy compares the adjacent frames directly in a shot, and the boundary frame of a shot are reserved. The two strategies are used for both query video filtering and the reference video filtering operations.2. We propose a consistency constraint based path verification algorithm. The algorithm verifies each path by employing the cross-clip accumulation sum of successive scores between query and reference frames. Experimental results show the proposed scheme can remarkably improve the localization precision.3. We propose a new CBCD algorithm, which is based on frame block fusion. Instead of fusing returned frames in frame level, the proposed method verifies the temporal consistency in the block level. The extensive experimental results show that the proposed scheme remarkably improves the detection accuracy of the video copy detection system in both localization accuracy and detection accuracy.