Feature selection based on human perception of image similarity for content based image retrieval

In this paper, we present an approach that enables selection of appropriate low-level features based on human perception for measuring image similarity in a Content-based image retrieval (CBIR) system. Human perceptual information is captured by three psychophysical experiments designed to explore different aspects of similarity. Their outcomes are used to formulate fitness functions in a genetic algorithm that selects a subset from an initial collection of popular low-level image features such as colour, texture and structure. The reduced subset of features is, thus, correlated to human perception and represents an attempt to bridge the semantic gap. We quantitatively validate our approach by building a CBIR system using such features and evaluating the retrieval precision.