Recovery of Illuminant and Surface Colors from Images Based on the CIE Daylight

Abstract

We propose a color constancy algorithm suitable for robot vision under natural environments based on the CIE daylight hypothesis. The algorithm can recover the illuminant color and the surface color in the scene from the R, G and B values observed on two color images, typically the current image and the past image. It utilizes the advantage of a robot which can exactly memorize images observed in the past. By employing the CIE daylight as a constraint, the stability and the accuracy of the color constancy algorithm based on multiple images, which we proposed previously, are remarkably improved. Effectiveness of the constraint is examined theoretically by analysing the behaviour of the algorithm under the existence of noise and also experimentally by using synthesized and real color images.