Vlad Cardei, Brian Funt, and Kobus Barnard, "Modeling color constancy with neural networks," Proceedings of the International Conference on Vision Recognition, Action: Neural Models of Mind and Machine, Boston, MA (1997).
The many algorithms used for color correction make a series of assumptions that try to constrain the problem of finding the scene illuminant under which a given image was taken. In contrast, the neural network we have developed has no explicit constraints. All rules are implicitly learned from the training set, which contains a large number of artificially generated scenes. The network estimates the chromaticity of the illuminant under which the given image was taken. This allows for a diagonal transformation of the image to another illuminant.
Full text (gzipped postscript)