Research On The Techniques For Content-based Video Copy Detection

With the rapid development of network technology and the fast growth of multimedia data, the problem of digital video’s illegal copy is getting serious. The exponential increases of illegally copied video have brought about serious crisis for digital copyright protection and harmful information filtering. Usually, the illegally copied video is processed with some kinds of copy attacks, which makes it not exactly the same to original video. However, their contents are almost the same. Therefore, content-based video copy detection is gradually becoming a hot research topic in the field of information security. It has great potential to be widely used in video copyright protection and video content management.After an analysis and survey of existing video copy detection algorithms, this thesis researches the techniques for video copy detection in both compressed domain and pixel domain, and two video copy detection schemes are proposed. The main works are summarized as follows.First, by fully utilizing the characteristics of compressed video, a video copy detection algorithm in compressed domain is proposed. It utilizes the ordinal measure of DC coefficients in the I frame of picture of group (GOP) to establish the index in video database, and the candidate-match video set are indexed by voting. First-stage matching is performed on candidate-match video set utilizing the mean value of modified DC coefficients in the blocks of I frame. The ordinal measure of energy is combined with Watson visual model for second-stage similarity-matching, and final copy detection results are obtained. Experimental results show that the proposed approach is robust for common copy attacks such as adding frame, luminance equalization and local cropping of image content, and significantly improves its detection performance without significant increase of detection time.Second, a robust hashing method based on speeded up robust feature (SURF) and ordinal measure (OM) is proposed for video copy detection. Since SURF is an invariant feature based on scale space theory, the local feature is extracted by SURF in a frame-by-frame manner. Then every frame is divided into4x4blocks, and every block is traversed by Hilbert-order rasterization to count the number of SURF points. The Hash value is built by the difference of SURF points in adjacent blocks in Hilbert curve. Two special copy attacks, i.e. picture-in-picture and video flipping, are discussed specifically. Extensive test experiments demonstrate the effectiveness of the proposed approach is improved in both accuracy and efficiency.