A decorrelating neural network for color constancy

Color perception remains roughly constant independent of the illuminant color. This is called color constancy, and much color research has focused on the color constant descriptor. In this paper, we describe a neural network model for color constancy according to the minimally redundant criterion proposed by Barlow. The proposed model with modifiable connections is trained by the anti-Hebbian rule to decorrelate the triplet of cone responses to 1569 Munsell color chips under three different illuminants. After the learning process, the network removes the effect of the change of illuminant by adapting to illuminant colors and achieves color constancy.