Luminance correlations define human sensitivity to contrast resolution in natural images

Luminance patterns encode shape and surface structure of objects in our environment. Humans can detect gradations of 1%-2% of background luminance. Is this level of sensitivity to luminance gradations (contrast) determined by the amount of ecologically meaningful information available in natural scenes? In the first experiment, subjects discriminated natural images I from their "posterized" versions I(n), in which the number of luminance gradations was reduced to n. In the second experiment, amplified residual images I(res)(n) infinity I - I(n) were discriminated from white-noise images, which lack any luminance correlations and thus information content. Performance in the two experiments matched remarkably well. Furthermore, as a function of n, the signal detected in both experiments was well fitted by the mutual information between nearby image pixels in the residual image I(res)(n). This suggests that human sensitivity to luminance contrast is optimized to extract ecologically useful information encoded by the luminance patterns of natural scenes.