Genetic algorithms for a family of image similarity models incorporated in the relevance feedback mechanism

Abstract

Abstract We propose two extensions of the Local Similarity Pattern/Weighted Local Similarity Pattern (LSP/WLSP) image similarity models, proposed by Stejic (2002). The objective is to improve the retrieval precision, by increasing the expressive power of the LSP/WLSP models. We formalize LSP and WLSP as special cases of a general similarity model, which defines image similarity as a weighted mean of the corresponding region similarities, and, in turn, each region similarity as a weighted mean of the corresponding feature similarities. In the case of LSP/WLSP models, region and feature weights have discrete non-negative values. The proposed two extensions are: (1) LSP-C+/WLSP-C+ models, with continuous non-negative weight values; and (2) LSP-C±/WLSP-C± models, with continuous positive and negative weight values. Similar to the LSP/WLSP, the proposed four models are incorporated in the relevance feedback mechanism, using genetic algorithm (GA) to infer the (sub-)optimal assignment of region and feature weights, which maximizes the similarity between the query image and the set of relevant images, chosen by the user. Accordingly, for the weight inference, GAs used for the LSP/WLSP models are extended from the integer-coded to the real-coded ones, and, in addition, new chromosome coding, two crossover modifications, six new mutation types, and a weights normalization operator are proposed. The proposed four models are evaluated on five test databases, with around 2500 images, covering 62 semantic categories. Compared with the existing image similarity models, including LSP/WLSP, and other models based on relevance feedback, proposed LSP-C±/WLSP-C± models bring in average over 10% increase in the retrieval precision. The main contributions, not limited to the LSP/WLSP models and the image retrieval, are: (1) the introduction of negative weights into the weighted mean of similarities, which increases the expressive power of the similarity model, and results in a significantly higher retrieval precision; and (2) new real-coded genetic operators suitable for the weight inference in general.