Occlusions and their relationship with the distribution of contrasts in natural images

An Information-Theory-like hypothesis recently proposed for early visual processing (the Minimal Local-Asperity hypothesis) accounts for the adaptive behavior with intensity of horizontal cells. It has been shown that for this to hold, the probability that a point is traversed by an occluding border must increase supralinearly (that is, with a positive second derivative) as a function of contrast. We test this condition by analyzing the distribution of contrasts and their relationship with occluding borders in natural images. We find that the distribution of contrasts in natural images falls exponentially as a function of contrast. Moreover, the probability that a point is traversed by an occluding border in natural images always rises with contrast until reaching one. This rise tends to be supralinear and addition of noise (at low intensities) increases the supralinearity, shifting the rising portion of the curve towards higher contrasts. These findings lend support to the Minimal Local-Asperity hypothesis, which proposes that one of the main roles of early retinal processing is to extract optimally edge, contrast, and luminance attributes from the visual world based on previous knowledge about natural images.