Design and Evaluation of a Content-Based Image Retrieval System

Abstract

The growth in size and accessibility of multimedia databases has changed our approach to information retrieval. Classical text-based systems show their limitations in the context of multimedia retrieval. In this chapter, we address the problem of conceiving and evaluating a content-based image retrieval system. First, we investigate the use of the query-by-example (QBE) paradigm as a base paradigm for the development of a content-based image retrieval system (CBIRS). We show that it should be considered as a complement to the classical textual-based paradigms. We then evaluate the capabilities of the most up-to-date computer vision techniques in contributing to the realisation of such a system. Further, beyond the necessity of accurate image understanding techniques, we show that the amount of data involved in the process of describing image content should also be considered as an important issue. This aspect of our study is largely based on the experience acquired by the text retrieval (TR) community, which we adapt to the context of CBIR. Similarly, the text retrieval community has also developed significant experience in evaluating retrieval systems, where judgements include subjectivity and context dependency. Extending this experience, we study a coherent framework for performing the evaluation of a CBIRS. As a practical example, we use our Viper CBIR system, using a novel communication protocol called MRML (Multimedia Retrieval Markup Language) to pinpoint the importance of the sharing of resources in facilitating the evaluation and therefore the development of CBIRS.