Research On Image Copy Detection Algorithms Based On The Bag-of-Words Model And Spatial Context Information

For copyright protection of digital images, the researchers have done a lot of research, Among which the content-based image copy detection has made great progress. Most of the image copy detection algorithms rely on the Bag-of-Words(BOW) model, which first quantifies the local features to be visual Words and then matches them. This model is able to effectively improve the matching efficiency.However, during the process of local feature quantification, the distinctive power of local features greatly reduced, leading to the emergence of variety of mismatches.And the mismatch may effects the copy detection accuracy. To solve the problem before, this paper proposes two image copy detection algorithms based on the combination of the BOW model and spatial context information.the specific research results are as follows:1) An image copy detection algorithm based on the combination of the BOW model and spatial context embeddingIn order to improve distinctive power of local descriptors, this paper proposes a novel spatial context embedding method.In the preprocessing phase,this scheme selects several stable contextual features for each SIFT feature, and then encodes the spatial relationship between it and the surrounding context features as spatial descriptors, finally embedding the descriptors in the index. In the process of copy detection, when each pair of SIFT matching is obtained by BOW quantification and matching, this pair of SIFT features will be further matched by spatial descriptors, to improve the distinctive power of local features. In addition, in order to improve the matching efficiency of spatial descriptors, we propose a secondary matching structure, which is able to effectively accelerate the matching process of spatial descriptor under the premise of reasonable matching effect.2) An image copy detection algorithm based on the combination of the BOW model and global context verificationAiming at the problem of low matching accuracy of the BOW model, the proposed scheme first obtains all the matching SIFT features between images by BOW quantification and matching, then uses of the rotation and scaling invariance of each pair of matching SIFT features to construct two overlapping areas, then gets two global context descriptor OR-GCDs from the overlapping areas. At last, the wrong SIFT matches are filtered by matching the global context descriptors. This scheme also puts forward an image similarity measurement method based on random verification, which can effectively improve the efficiency of copy detection by avoiding verifying all matches. In addition, we also further extend our algorithm to make it feasible to solve the problem of partial copy image detection, by locating the potential copy image area before copy detection.