Abstract

The image of a natural scene can be considered a product of reflectance and effective illumination functions. One goal of image understanding is the estimation of the reflectance and illumination from image luminance data. Human observers are remarkably good at inferring reflectance changes from an image. One well-known aspect of this ability is lightness constancy. However, because there are two unknowns for every data point in the image, this problem is ill-posed, and it does not have a unique solution without prior constraints on the class of reflectance and illumination functions. We use Markov random fields to model and thereby constrain the class of reflectance and illumination functions. Our computational goal is taken to be the maximization of the posterior distribution of the reflectance and illumination conditional on the image. This goal is achieved using stochastic relaxation. This approach provides a general framework for quantifying the computational theory apart from specific algorithmic implementations for a number of problems of early vision.

© 1987 Optical Society of America