Image retrieval and perceptual similarity

Abstract

Simple, low-level visual features are extensively used for content-based image retrieval. Our goal was to evaluate an image-indexing system based on some of the known properties of the early stages of human vision. We quantitatively measured the relationship between the similarity order induced by the indexes and perceived similarity. In contrast to previous evaluation approaches, we objectively measured similarity both for the few best-matching images and also for relatively distinct images. The results show that, to a large degree, the rank orders induced by the indexes predict the perceived similarity between images. The highest index concordance employing a single index was obtained using the chromaticity histogram. Combining different information sources substantially improved the correspondence with the observers. We conclude that image-indexing systems can provide useful measures for perceptual image similarity. The methods presented here can be used to evaluate and compare different image-retrieval systems.