Interactive Image Retrieval by Means of Abductive Inference

Current research and development in image retrieval has yielded an extensive collection of feature extraction algorithms. Just to mention a few, texture analysis, sketch comparison, color similarity measures, etc. have shown to be useful matching tools in certain domains. Unfortunately, the design of current image retrieval systems tends to be somewhat one-sided, stressing mainly technical issues. This paper describes a formal framework to provide means for effective, interactive retrieval sessions. The key idea is to model the characteristics of image retrieval algorithms. Each algorithm is described in terms of qualitative rules which reflect its average performance in a manually classified training set of randomly selected images. The abductive retrieval system MIRACLE makes uses of the rules to map query statements to the appropriate set of image retrieval operators. Users obtain qualitative feedback on how to optimize their query statements with respect to a given information need and to the power of the image feature extraction algorithms.