12. Understanding color categories, color constancy, color induction, and lightness perception from information theory

Li Zhaoping z.li@ucl.ac.uk

Computer Science, UCL

I explore an understanding of colour appearance predicated on the brain’s mapping sensory inputs into discrete categories conveying the maximum bits of Shannon information about the input. Under sufficiently high (but not infinite) signal-to-noise ratio, when an input ensemble contains the usually large dynamic range, an information maximizing mapping from the contrast-gain-controlled photoreceptor inputs to, e.g., six, categories typically carves the input space into regions that correspond to the perception of white, black, red, green, blue, and yellow colour categories. This input-to-category mapping corresponds to another mapping from surface reflectance to category of colour appearance. Illumination changes that sufficiently preserve signal-to-noise can alter the input-to-category mapping but leave the reflectance-to-category mapping almost unchanged, achieving colour constancy. This hypothesis of informationally optimal colour boundaries, when applied to small input ensembles made of inputs from a single or a part of a scene, can account for various colour illusions in particular color induction, and, under achromatic inputs, typical phenomena in lightness perception. It does not at present accommodate the spatial configuration factors that influence colour/lightness appearance.