Robust dichromatic colour constancy

Abstract

A novel colour constancy algorithm that utilises both physical and statistical knowledge is introduced. A physics-based model of image formation is combined with a statistics-based constraint on the possible scene illumination. Based on the dichromatic reflection model, the intersection of two colour signal planes from two objects will yield the scene illuminant. However, due to noise and insufficient segmentation this approach tends not to work outside the lab. By removing those intersections that are likely to produce unstable illuminant estimates and applying an illumination constraint in form of a set of feasible reference lights the colour constancy algorithm presented clearly outperforms the conventional approach and provides excellent results on a benchmark set of real images. © Springer-Verlag 2004.