Relevance feedback for content-based image retrieval using proximal support vector machine

In this paper, we present a novel relevance feedback algorithm for content-based image retrieval using the PSVM (Proximal Support Vector Machine). The PSVM seeks to find the optimal separating hyperplane by "regularized least squares". The obtained hyperplane comprises the positive and negative "proximal planes". We interpret the proximal vectors on the proximal planes as the representatives among training samples, and propose to use the distance from the positive proximal plane as a measure of image dissimilarity. In order to reduce computational time for relevance feedback, we introduce the "expanded sets" derived from the pre-computed dissimilarity matrix, and apply the feedback algorithm to these expanded sets rather than the entire image database, while preserving the comparable precision rate. We demonstrate the efficacy of the proposed scheme using unconstrained image databases that were obtained from the Web.