Multiple Strategies for Relevance Feedback in Image Retrieval

Retrieval by similarity in image databases can be accomplished by comparing features extracted from a sample image to the features extracted from the images in the database. Relevance feedback is a technique of query refinement used in text as well as in image retrieval systems which is based on user selection of the most relevant images among the ones retrieved, whose features are used to modify the original query, further resubmitted to the retrieval system. This paper presents an approach to relevance feedback analysis based on three different strategies which can be combined. The first strategy considers the discrimination power of the feature values with respect to the image population. This technique, based on variance analysis, is used in text retrieval systems too. The second strategy compares several feature computation techniques, choosing the one that best matches the user selection of relevant answers. The third technique considers the discrimination power of the different features, enhancing the contribution of the features that best identify the relevant images marked by the user. The three techniques can be combined into a global evaluation, giving better retrieval responses. The techniques have been implemented in a retrieval system that computes image similarity based on color, texture and shape analysis according to several algorithms. The user can select several metrics or correlation coefficients to compute image similarity. A discussion of some design issues and an evaluation of the relevance feedback strategies on a sample image database is discussed.