Research And Implementation On System Of Ordinal Measure Based Video Copy Detection

With the development of digital technology and network multimedia, it’s easy to make, broadcast, and distribute videos. Digital videos are always reformatted to meet different applications. Copies make the management of growing video data be a significant problem, for example, video copyright protection and efficiency of video retrieval. Existing physical protection and digital watermark can only be operated on videos undistributed. Relatively, content-based video copy detection can detect videos not only undistributed but also distributed, but its performance is to be improved. This thesis studies how to improve the performance.Traditional detection methods extract part of all frames, use the feature of frames to construct signatures, and judge the similarity of two videos by matching the frame sequence. Now, copies are made by reformatting and editing source videos several times, and broadcasted widely. Some noise data are added into video data while being edited, so that traditional methods can’t perform well like before. Otherwise, frames extracted by traditional methods always have some redundancy which increases the size and influences the result of sequence matching.Based on the study of ordinal measure feature, the thesis proposes a new method to extract frames and two methods to eliminate redundancies to resolve the problem of frame redundancy. At the same time, ordinal measure feature is used to match frames to resolve sensitive feature matching. Experimental results show that our method performs well in some common copy detections.The thesis focuses on three aspects. First, this thesis summarizes and analyzes methods of selecting frame sets, studies the feature of redundant data, and proposes a new method to extract frame sets based on illumination abruption. At the same time, the thesis proposes two methods to filter Key Frames. One uses ordinal measure to eliminate redundancy spatially, and the other one uses shot length to eliminate temporally. Secondly, based on the analysis of feature-selected sensitive in sequence matching, the thesis selects-ordinal measure feature as the sequence-match feature, and proposes a loose computation scheme of feature distance to increase the robustness of the sequence matching. At last, in allusion to two aspects, a new similarity computation scheme is proposed.Finally, a prototype system has been implemented, and some experiments has been done. Consequently, the thesis compares and analyzes the result. The results prove the copy detection solution proposed in the thesis is superior and effective.