An efficient core-area detection algorithm for fast noise-free image query processing

Recent content-based image retrieval techniques enable users to arbitrarily exclude noise (i.e. irrelevan t regions) from image similarity consideration. This capabilit yhas resulted in high retrieval effectiv eness for a wide range of queries. T o support large image collections, subimages of a predetermined base shape (e.g., circle, polygon) are collected and indexed into a m ultidimensionalaccess structure. A tthe query time, an area of such a shape enclosing part of the queried objects, called the core area, will be identified and used in the initial search of potential candidates before an appropriate detailed similarity measure is performed on the original query. Iden tifying the core area of a query can be challenging as it is allow ed to con tain certain noise and ma y not be unique. In this paper, w e propose an efficient algorithm, called the Seed-Gr owingDete ctionA lgorithm, to automatically detect the optimal core area. We have implemented the proposed technique in our image retrieval system for a large database. Our experimental results sho w that our approach is effective and able to minimize time overhead of query preprocessing.