Image Retrieval by Regions: Coarse Segmentation and Fine Color Description

Abstract

In Content-Based Image Retrieval systems, region-based queries allow more precise search than global ones. The user can retrieve similar regions of interest regardless their background in images. The definition of regions in thousands of generic images is a difficult key point, since it should not need user interaction for each image, and nevertheless be as close as possible to regions of interest (to the user). In this paper we first propose a new technique of unsupervised coarse detection of regions which improves their visual specificity. The Competitive Agglomeration (CA) classification algorithm, which has the advantage to automatically determine the optimal number of classes, is used.

The second key point is the region description which must be finer for regions than for images. We present a novel region descriptor of fine color variability: the Adaptive Distribution of Color Shades. It is based on color shades adaptively determined for each region at a high resolution: 5 million of potential different colors represented against few hundreds of predefined colors in existing descriptors.

Successful results of segmentation and region queries are presented on a database of 2500 generic images involving landscapes, people, objects, architecture, flora. . . .