Abstract
How the visual system encodes light patterns on the retina and transforms this information into the full range of perceptual values that humans experience is arguably the central challenge of vision. A key aspect of this problem is how such encoding generates perceptions of brightness. A range of observations dating back a century or more have shown that the same amount of light can elicit radically different perceptions of target brightness as a function of context. Debate about the coding strategy responsible for these phenomena, initiated by Helmholtz, Hering, Mach and others, continues today. One approach to understanding coding has been to assume that the neural code for luminance that ultimately elicits brightness percepts arises from low-, intermediate- and high-level neural processing. This assumption, however, has not led to a satisfactory account of the percepts elicited by a variety of stimuli.
Based on an extensive analysis of natural scenes, we suggest that the encoding of luminance is governed by the probability distributions of luminance values in such stimuli, the rationale being a means of dealing efficiently with all possible occurrences of luminance patterns in typical images. Such coding thus represents the conditional probability distribution of target luminance within contextual light patterns, always making use of the full capacity of the system. The brightness of any target would then correspond to the value of the underlying luminance encoded in this way. In confirmation of this idea, the relevant probability distributions obtained from natural images predict a wide range of otherwise difficult to explain brightness phenomena. These results support the conclusion that the visual system encodes luminance according to the probability distributions of the co-occurring luminance values experienced in natural environments, and that the ensuing brightness percepts are always a consequence of this optimal coding strategy.