Abstract

The spectrum of the light from a scene depends on its reflecting properties. Differences in reflectances allow surfaces to be discriminated and, potentially, uniquely identified. But inferences about the content of a scene cannot be made directly from the reflected light. This is because reflected spectra from the same surfaces change with changes in scene illumination; and, conversely, under the same illumination, reflected spectra from different surfaces can produce the same responses in the receptors of the eye. In short, there is no one-to-one correspondence between surfaces and receptoral activity. Given these uncertainties, what are the physical limits on an observer’s ability to identify surfaces, especially in natural scenes? The aim here is to estimate how much information in a scene is actually available to an ideal trichromatic observer; how efficiently this information is coded at receptoral and post-receptoral levels; and how efficiently it can be retrieved in practice in a simple task requiring point-by-point identification across images of a scene. It is shown that coding at receptoral levels is highly redundant, but postreceptorally, its efficiency increases markedly, by about an order of magnitude. Critically, a postreceptoral transformation providing minimum redundancy in the information available yields the maximum information retrievable in identification.