Research On Local Feature Based Image Copy Detection

With the development of Internet and multi-media technologies, the distribution, spreading and modification of multi-media contents becomes more convenient and at the same time, a challenge is posed to the copyright protection. As one of the methods of copyright protection ,content-based copy detection technology receives extensive research and applications.Content-based image copy detection technology mainly includes two parts: feature extraction and index construction, which are also the research hot spots and difficulties in content-based image copy detection. Although local features can resist certain geometric attacks like rotation, scale and transformation etc, the algorithms is complicated and inefficient in execution; as for index construction, current index technologies often encounter with dimension curse, although many improved methods are proposed, the results are not remarkable.Basing on reading the literature at home and abroad, systematic research on content-based copy detection is conducted and image copy detection basing on SIFT and LSH is implemented and then,multi-probe LSH aimed to improve the basic LSH which employs too many hash tables and too much storage space is introduced .The experiment uses the benchmarks precision-recall and the query time to assess the system, the result shows both the precision and the recall ratio reach above 90% with less hash tables that is one ninth the original one's ,but the mean query time has delayed for some extent.In the domain of text retrieval, inverted index technology is especially suitable to query in large scale dataset with high effiency.Therefore, inverted index is introduced into image copy detection. Experiment result shows both the detection quality and query time are superior to LSH and multi-probe LSH.