Skip to Main Content
The quality of codebook is the determinant factor in BoW-based copy detection strategies. However, most of the adopted codebook construction algorithms are derived from image retrieval or object recognition, which neglect the robustness in partitioning original features and copy features (especially those with serious transformations) into the same group. To deal with this problem, we have developed an Incremental Clustering algorithm to construct a robust codebook. Unlike many existing algorithms which need loading the entire data into memory, our algorithm process data incrementally and improve the quality by transferring data from a global view. In addition, we propose a codebook evaluation scheme by simulating copy and non-copy pairs. Our experimental results show that our approach attains a high precision in copy pairs, which demonstrates the robustness of our codebook.