Indexing, learning and content-based retrieval for special purpose image databases

This chapter deals with content-based image retrieval in special purpose image databases. As image data is amassed ever more effortlessly, building efficient systems for searching and browsing of image databases becomes increasingly urgent. We provide an overview of the current state-of-the art by taking a tour along the entire "image retrieval chain"-from processing raw image data, through various methods of machine learning, to the interactive presentation of query results. As it often constitutes the key to building successful retrieval systems, we first discuss the issue of content representation and indexing. Here both the computation of global and local characteristics based on image segmentations is reviewed in some detail in the context of interior design images. Also the representation of content by means of MPEG-7 standard metadata is introduced. In regard to the search system itself, we focus particularly on interfaces and learning algorithms which facilitate relevance feedback, i.e., on systems that allow for natural interaction with the user in refining queries by means of feedback directly in terms of example images. To this end the literature on this subject is reviewed, and an outline is provided of the special structure of the relevance feedback learning problem. Finally we present a probabilistic approach to relevance feedback that addresses this special structure. © 2005 Elsevier Inc. All rights reserved.