Color constancy from physical principles

A well known property of human vision, known as color constancy, is the ability to correct for color deviations caused by a difference in illumination. A common approach to investigate color constant behavior is by psychophysical experiments, regarding the human visual system as a black box responding to a well defined change in an laboratory setup. A fundamental problem in psychophysical experiments is that significant conclusions are hard to draw due to the complex experimental environment necessary to examine color constancy. An alternative approach to reveal the mechanisms involved in color constancy is by modeling the physical process of spectral image formation. In this paper, we aim at a physical basis for color constancy rather than a psychophysical one. By considering spatial and spectral derivatives of the Lambertian image formation model, object reflectance properties are derived independent of the spectral energy distribution of the illuminant. Gaussian spectral and spatial probes are used to estimate the proposed differential invariant. Knowledge about the spectral power distribution of the illuminant is not required for the proposed invariant. The physical approach to color constancy offered in the paper confirms relational color constancy as a first step in color constant vision systems. Hence, low-level mechanisms such as color constant edge detection may play an important role in front-end vision. The research presented raises the question of whether the illuminant is estimated at all in pre-attentive vision. (C) 2002 Elsevier Science B.V. All rights reserved.