Approximate Query Processing for a Content-Based Image Retrieval Method

Abstract

An approximate query processing approach for a content-based image retrieval method based on probabilistic relaxation labeling is proposed. The novelty lies in the inclusion of a filtering mechanism based on a quasi lower bound on distance in the vector space that effectively spares the matching between the query and a number of database images from going through the expensive step of iterative updating the labeling probabilities. This resembles the two-step filter-and-refine query processing approach that has been applied to k-nearest neighbor (k-NN) retrieval in database research. It is confirmed by experiments that the proposed approach consistently returns a "close approximation" of the accurate result, in the sense of the first k' in the top k output of a k-NN search, while simultaneously reduces the amount of processing required.