Construction-and-extraction based index for images retrieval

The past recent years have witnessed more and more applications on image retrieval. As searching a large image database is often costly, to improve the efficiency, high dimensional indexes may help. This paper proposes an adaptive hybrid index (AHI) supported by a construction-and-extraction technique to support image retrieval. First, the image clusters are further partitioned into sub-clusters to reduce the overlap between clusters and indexed into an iDistance index. Then, the query sampling statistically extracts some sub-cluster from the iDistance index into a sequential file. Finally, the users' queries are accurately returned by searching both the iDistance index and the sequential file. It's proved that the proposed AHI never performs worse than the sequential scan. Particularly, the experimental results demonstrate that the proposed index AHI is beneficial and achieves better performance than some exiting methods. It is about 2 times faster than iDistance, almost three times than Omni-sequential, more than four times faster than sequential file and more than 10 times faster than M-tree on the benchmark images set. The effect of the proposed AHI is also investigated by our implemented content based images retrieval system.