Content-based retrieval from image databases: Colour, compression, and browsing

Content-based image retrieval (CBIR) has been an active research area for the last two decades and much progress has been made in that time. However, there are still many challenges to be overcome and in this paper we highlight some of these together with some approaches that we have developed to address these problems. In particular, we focus on the issues of colour (or more precise colour variance), image compression, and image database browsing. While colour features are the most widely used image descriptors for CBIR, colour is not necessarily a stable cue as it also depends on various image capture conditions. Colour invariants are features designed to be robust with respect to these confounding factors. Image compression, which is typically applied to most images in use, leads to both processing overheads but also to a small but noticeable drop in retrieval performance. To address these problems, we have developed image retrieval techniques that operate directly in the compressed domain, yet provide better retrieval performance than many standard CBIR techniques. Finally, we look at browsing systems as an alternative approach to dealing with large image databases. The hue sphere browsing system that we have developed organises images on a spherical visualisation space so that visually similar images are located close to each other, and provides powerful interaction tools that aid in exploring the visualised image collections.