Computational Colour Constancy by using two learning machines: Contributions to Neural Networks and Ridge Regression for illuminant estimation

Abstract

Many applications of Colour Image Processing have problems with changes of the illuminant, in this work we present some contributions to the research line called Computational Colour Constancy that could provide solutions to those problems. Colour Constancy is the ability through which colour perception remains almost constant despite changes in the illuminant and Computational Colour Constancy can be defined as the emulation of this biological phenomenon through computer programs. There are a lot of paths to perform this task and one set of these propose to estimate the illuminant by learning machines before discounting the illuminant. In this work we present the results of some improvements to Computational Colour Constancy by using Neural Networks and Ridge Regression approach to Colour Constancy. The main contribution is the extension of the original algorithms to a 3-D spectral spaces and additionally a cooperative approach between these two learning machines is commented and a stricter evaluation to Ridge Regression was performed. The experiments with a combination of synthetic data and real images showed the advantages of our proposal over some classical methods for illuminant estimation.