Learning and adaptive characterization of visual contents in image retrieval systems

Accurate characterization of visual information is an important requirement for constructing an effective content-based image retrieval (CBIR) system. Conventional nonadaptive image models adopted in simple CBIR systems do not usually perform this task adequately. A more effective method is to adopt a machine learning approach where the system directly learns the image interpretation criteria from the users. As a result, we adopt neural network techniques to perform this characterization. More specifically, we have adopted a radial basis function (RBF) network for implementing an adaptive metric in image retrieval which progressively models the notion of image similarity through continual relevance feedback from users and a modular neural network for characterizing edge information in the front-end feature extraction module of the CBIR system. Experimental results have shown that the proposed methods outperform conventional techniques in terms of both accuracy and robustness.