Color-Calibration of Robot Vision System Using Self-Organizing Feature Maps

This paper presents an application of Kohonen's self-organizing feature maps (SOM) for solving the problem of color constancy. The main problem is to evaluate the transformation between collections of color-points forming diierently shaped clouds in color space under changing illumination. The main idea is to embed appropriate 3D coordinate systems into these clouds by self-organization, and so to be able to nd corresponding color points within diierent clouds. The diierence of the locations of corresponding neurons in two SOMs is then an approximation for the particular color shift belonging to the diierence between a reference illumination and a given illumination. The observed shifts provide a table of vectors in the color space, which in correction steps can be applied to color images taken under a given illuminant.