Research On Content-similarity Based Video Segment Copy Detection

The rapid development of Internet technologies and continuous proliferation of video editing tools lead to exponential increase of illegal video copies.Especially,there are huge amounts of near-duplicate videos over networks.Content-based video copy detection has been becoming one of the hottest topics in the field of video copyright protection and content management.In this thesis,the existing techniques for video copy detection are summarized,and emphasis is put on the detection for those short near-duplicate video clips.Its main difficulties arise from the uncertainty of copy length and copy location.Therefore,the main motivation will be how to rapidly detect and locate those similar clips from the candidate video.The key techniques involved are the extraction of robust,compact and discriminative features,the efficient similarity search and accurate video sub-sequence matching.After summarizing existing work about these three key techniques,we make some active attempts to address these issues.The main work and contributions are summarized as follows:First,an inverted index method is proposed for the feature indexes of video sub-sequence.The features of single frame is obtained by ordinal measure(OM)combined with local binary pattern(LBP).They are combined with spatio-temporal information and encoded to form independent feature vectors.The inverted index is used to map the video sequence tapped specific ID number into the corresponding index cell,combined with the order consistency principle to search valid copy clips.Experiments show that the approach can effectively locate one or more similar clips in the query video simultaneously,and reduces the time complexity of matching thereby improve the detection efficiency.Second,a video copy detection approach is proposed using robust SIFT-based hash feature and two-stage matching.The proposed robust hashing feature is built on the basis of SIFT feature points,which contains the information of circular blocks and spatial distribution of interest points.It reduces the feature dimension and computation complexity,and is resilient to geometric attacks such as rotation and zooming.To accurately locate the similar clip from candidate video,a two-stage matching method is adopted.Key frames are used to represent video segment,and the similarities of video segments are used to mark the matching of segments.Graph-based approach is used to find the longest match path among matching segments,thus locate similar clips.Experimental results prove the robustness of the proposed feature when dealing with a variety of attacks,and also verify that the proposed two-stage matching scheme can obtain more precise location of video segment in query video.