A fast refined search method based on filtering by query difference

This paper proposes a novel method for fast refined search based on the relevance feedback in various content based multimedia retrieval systems. The basic idea of the proposed method is filtering of the irrelevant multimedia objects using the intermediate results at the first search and the difference between the original and the refined query. The meaning of the intermediate results is the distances to the original query for all objects in a database. For each object, the approximation of distance to the refined query can be simply calculated by subtracting the query difference from the distance to the original query. This mechanism helps to reduce both CPU and I/O time of the refined search because many irrelevant objects can be filtered out by the approximations of distance which are calculated by only one operation without reading the high dimensional objects. Also, it can be used with the filtering based indexing methods which are developed to resolve the problem called "curse of dimensionality". Upon experimental results, the refined search using the proposed method is about 5 times faster than the simple sequential search.