Fast Image Retrieval System Using Representative Images

Image retrieval is common behavior due to the dramatically increasing of image contents. From the point of view of service provider, image retrieval service is considered as a second option because if image database size is getting larger, then it seems to have to purchase a new machine. To solve this problem many reasonable solutions had been proposed. Typical solution is clustering to reduce the search space. In this paper, we suggest another approach using representative images which are similar to cluster heads. We experiment our approach with 1.6 million images using MPEG-7 color structure descriptor which is one of the most powerful descriptor to retrieval the relevant color images. Finally we show that our approach is at least 85% similar to exhaustive search results in top 10,000 while supporting response time is less than 0.2 s. And we expect that our approach may solve the scalability regardless of image database size.