The Bright-chromagenic Algorithm for Illuminant Estimation

In this paper, we propose a new algorithm for illumi-nant estimation. We begin by reviewing the concept of chro-magenic colour constancy, where two pictures are taken from each scene: a normal one and one where a coloured filter is placed in front of the camera, and look at param-eters known to affect its performance such as filters and sensor choice. We show that the basic formulation of the chromagenic algorithm has inherent weaknesses: a need for perfectly reg-istered images and occasional large errors in illuminant es-timation. Our first contribution is to analyse the algorithm performance with respect to the reflectances present in a scene and demonstrate that fairly bright and desaturated reflectances (e.g., achromatic and pastel colours) provide significantly better chromagenic illuminant estimation. This analysis leads to the bright-chromagenic algo-rithm. We show that it not only remedies the large error problem but also allows us to relax the image registration constraint. Experiments performed on a variety of syn-thetic and real data show that the newly designed bright-chromagenic algorithm significantly -in a strict statistical sense- outperforms current illuminant estimation methods, including those having a substantially higher complexity.