Abstract
Two rapid mechanisms in the early visual system control response gain on the basis of locally prevalent luminance and contrast. We asked whether the arrangement and operation of these gain control mechanisms are matched to the statistics of the environment. First, we measured local luminance and local contrast in random patches of calibrated, 12-bit gray-scale natural images (van Hateren and van der Schaaf, 1998). These patches correspond to the images that would fall on a receptive field during successive fixations. We found that luminance and contrast are statistically independent of each other (correlation −0.2). This remarkable independence is a property of real world images, but not of artificial images: In phase-scrambled images, which have the same power spectra as natural images, luminance and contrast are strongly correlated (correlation −0.8). Second, we characterized mechanisms of luminance gain control and contrast gain control in the lateral geniculate nucleus (LGN) of anesthetized and paralyzed cats. Stimuli were temporal frequency sweeps (gratings whose temporal frequency varied exponentially from 0.5 to 40Hz over 5s). From the responses we estimated one impulse response for each of 12–25 combinations of luminance and contrast covering ranges encountered during natural viewing. The estimated impulse responses explain 85% (median) of the variance in the responses to sweeps. Imposing separability, i.e. that luminance gain control has the same effects at all contrasts, and contrast gain control has the same effects at all luminances, has little effect on the predictions: the best separable set of impulse responses still explains 81% of the variance. Thus, functionally speaking, luminance and contrast gain control operate independently, reflecting the very independence encountered in natural images.
Supported by James S McDonnell Foundation and NIH grant R01EY11747