Color recipe prediction by artificial neural networks

The conventional 2-flux theory of Kubelka and Munk employed for computer colorant formulation reaches its limits in certain areas of coloration suggesting to look at an alternative approach. Improved results of the model can only be achieved by detouring deeply into radiative transfer theory. An alternative approach avoiding such mathematical complexities is offered by recent developments in the field of artificial neural networks. By adopting a multilayer feedforward network model several nets of different topologies have been trained by means of the Kubelka/Munk equation for reflexion of opaque diffusers and experimentally determined optical material parameters of a vendible automotive repairing system. In contrast to previous investigations calculations have been performed on reflexion and not on color space, and the number of colorants have been extended considerably. Elaborate analysis of the results utilizing different strategies for pattern generation clearly demonstrates that f eedforward networks training error deteriorates considerably with increasing number of colorant components (n > 5).