<title>Image retrieval and reversible illumination normalization</title>

We propose a novel approach to retrieve similar images from image databases that works in the presence of significant illumination variations. The most common method to compensate for illumination changes is to perform color normalization. The existing approaches to color normalization tend to destroy image content in that they map distinct color values to identical color values in the transformed color space. From the mathematical point of view, the normalization transformation is not reversible. In this paper we propose to use a reversible illumination normalization transformation. Thus, we are able to compensate for illumination changes without any reduction of content information. Since natural illumination changes affect different parts of images in different amounts, we apply our transformation locally to sub-images. Basic idea is to divide an image into sub-images, normalize each one separately, and then project it to an n-dimensional reduced space using principal component analysis. This process yields a normalized texture representation as a set of n-vectors. Finding similar images is now reduced to computing distances between sets of n-vectors. Results were compared with a leading image retrieval system.