Non-metric similarity ranking for image retrieval

Abstract

Over many years, almost all research work in the content-based image retrieval has used Minkowski distance (or Lp-norm) to measure similarity between images. However such functions cannot adequately capture the aspects of the characteristics of the human visual system. In this paper, we present a new similarity measure reflecting the nonlinearity of human perception. Based on this measure, we develop a similarity ranking algorithm for effective image retrieval. This algorithm exploits the inherent cluster structure revealed by an image dataset. Our method yields encouraging experimental results on a real image database and demonstrates its effectiveness. © Springer-Verlag Berlin Heidelberg 2006.