Visual Information Retrieval

1. Introduction Present methods of retrieving images based on visual content largely center around the extraction of certain salient features of an image, such as color, texture, shape, and the computation of a similarity measure between two images. Such methods, while effective in particularly focused applications, fail to generalize because they are semantically primitive when compared to human similarity judgment. Furthermore, most methods fail to exploit domain knowledge and relevance feedback to improve the accuracy of the search. These techniques, although commonly practiced in text-based retrieval systems, are lacking in content-based systems. Clearly, there is much more to be done in this area; we need to move towards a more intuitive, human-like query facility in order to allow laymen to use multimedia databases. In this article, we discuss the development of a multi-layered framework for visual information retrieval. In this framework, we integrate current works on content based information retrieval at various level of abstractions – as pixels, as features, as objects, or as abstract concepts. Retrieval may be carried out at different layer. Relevance feedback may also be used at different layer to improve retrieval performance. This paper describes the design, implementation and testing of the new system.