Abstract

Let image intensity be measured through each of three spectral filters S, M, and L. Let each image region be mapped to a point in a color space R. If the illuminant is white, points in different quadrants of R will correspond to distinct materials.¹ Suppose that given a random collection of materials, half reflect more light in the S sample than in the M, and independently, half reflect more in the L than in the M. (Any fixed percentages will do.) Then an image under white light of a random set of materials will, when mapped into R, yield approximately the same number of points in each quadrant. If an image of random materials maps into R so that quadrants are unevenly populated, it can be inferred that the illuminant is not white. Normalization coefficients s and I can then be found so that the quadrants of the new space are equally populated. One crucial problem is that many natural images (e.g., forest scenes) do not contain a random collection of materials. The spectral crosspoint operator² can specify a subset of image regions more likely to represent a random set of materials than the set of all image regions (which might contain repeated materials). Normalization can then proceed with this subset.

© 1986 Optical Society of America