A novel method for long term learning using cluster merging

Abstract

A rudimentary problem in CBIR is the semantic gap that exists between the low-level features and high-level semantics as perceived by humans. Relevance Feedback is one solution to bridge this gap. Usually relevant images corresponding to a query form many islands in feature space. Also a query has set of relevant images, and vice-versa an image has many queries. This principle is used in formulating our novel method of Long Term Learning(LTL). Here the islands of relevant images corresponding to a query are tagged with query index. These islands of many queries are merged based on some criteria like overlapping, nearness etc., to obtain overall merged clusters with associated tagged query indices. Given a query, the initial retrieval set is obtained by combination of two schemes, one uses RF log that covers search based on nearby concept images and the other is based on simply the Mahalanobis distance of the query with database images. Experimental results illustrate the potential of proposed method of LTL by cluster merging in initial and as well over the iterations. The efficiency of our method is shown using standard Wang database of 1000 images.