Large-Scale Image Copy Detection Algorithm Research Based On Local Media-Fingerprints

With the rapid development of Internet and multimedia technology, multimedia iseasier to be spread, copied, and tampered with. However, the multimedia copyrightprotection has become an important issue. Content-based copy detection, which is aneffective technique to solve this problem, includes feature extracting and indexing system.Nowadays, the difficulties of the Content-based copy detection lie in how to increase thefeature’s robustness and decrease the retrieval time.In feature detection, SIFT detector outperforms state-of-the-art approaches, but it stillhas some limitations. Inspired by multi-resolution histogram, we propose Multi-scale SIFTdescriptor to improve the feature’s discriminability. Based on SIFT descriptor, SURFdescriptor, Multi-scale SIFT descriptor, and with the assitance of PCA and binary code, weget three local media-fingerprints. Experiments show that the Multi-scale SIFTmedia-fingerprint performs best, because it is more discriminative and more robustnessthan the other two descriptors.As to the indexing structure, we implement an efficient inverted indexing system,which consists of lookup tables to speed up the efficiency. Thus, the system performancesinterms of robustness and efficiency have been improved. We implement theMedia-fingerprint simility search based on hamming distance. This method couldeffectively resist media-fingerprints’ variation after being attacked, and enhance thefeatures’ robustness. We implement distributed indexing to speed up the average responsetime, and use lookup sub-table and hash-based function for fast location of index.The experiments are done on a collection of a million images prove that Multi-scaleSIFT media-fingerprint performs the best, which verifies the robustness of the Multi-scaleSIFT descriptor. In addition, we implement the efficient inversed indexing system, andimprove the system performance in two different aspects.