A novel method for image retrieval using relevance feedback and unsupervised clustering

Abstract

The standard approach to content-based image retrieval is currently concerned with bridging the semantic gap or the gap between the results produced by the use of low-level features and the human end-user expectations based on high-level semantics. In this paper, we suggest that there are advantages to bridging the gap in two stages by proposing an intermediate level. We show that unsupervised clustering of low-level image features provides a suitable basis for an intermediate level representation and define a CBIR system using such an approach. The main advantages of using an intermediate level are (a) it is not necessary for all positive responses to a user query be categorized into a single class; (b) it is possible to overcome the small-sample problem with too few positive examples; and, (c) to improve performance without greatly increased computational cost. Experimental results on Wang's database (1000 images) and Corel Photo gallery (10,800 images) show that the intermediate level analysis leads to better results. Copyright 2011 ACM.