Multiple Feature Relevance Feedback in Content- Based Image Retrieval using Probabilistic Inference Networks

Abstract

The BIR content based image retrieval system uses a Bayesian belief network architecture to match query by example images to images in a database. This probabilistic architecture provides support for multiple image features at varying levels of abstraction. Relevance feedback may be natively implemented in the model via diagnostic inference in the Bayesian network. In this paper we describe how different relevance feedback scenarios can be dealt with in this system, in particular those involving feedback involving multiple classes of image features. In addition, a feature weighting scheme is proposed in order to automatically apportion flows of diagnostic inference according to the relative importance of visual features in the images chosen as relevant by the user.