Introduction

The color of an object in the image is widely used in query by image content systems (e.g. [3]). The popularity of color as an index resides in its ease of computation. However, if the same scene is captured under different conditions the color of the objects may change significantly. Beside the well known dependence on the spectral composition of the illuminant (e.g., the recorded colors change significantly when the illumination changes from daylight to indoor incandescent light), an image can also appear significantly different when recorded in differently calibrated systems (photographic and/or electronic).

Above two color images of the same scene are shown, scanned in from [4, pp. 90-91]. The normal lens image was taken from 50m while the telephoto lens image was taken with a telephoto lens from 1000m. In the latter all the colors appear less saturated and the gray tones dominate. Human observers have no difficulty in recognizing the two images as being the same, while a simple color based query would fail to match them.

Compensating for the effect of an illuminant is known as the color constancy problem. See for example, [9] for a review of the major techniques. The goal of these techniques is to infer about the reflectance properties underlying a given image, which is more than what is required for the comparison of two similar images. In the latter case the differences can be captured by defining a relative illuminant, i.e., the homogeneous illumination which applied to one of the images will transform its chromaticity distribution into that of the other image. In this paper we propose a simple, statistical approach to estimate the relative illuminant. The approach exploits spectral information which has not been widely used so far in computer vision research.