Abstract
The distribution of contrast in natural scenes is highly non-uniform. We examined how this variability affects apparent contrast and contrast discrimination thresholds. Natural images were randomly selected from the Van Hateren image database and viewed through one hundred non-overlapping pseudo-randomly positioned Gaussian apertures. The local rms contrast of each aperture was systematically manipulated, thus allowing control of contrast variation and mean contrast across the image. Observers performed two main tasks, first a matching task in which they adjusted the contrast of an image in which all apertures had the same rms contrast to match that of an image whose contrast was varied across the image but otherwise was identical. In the second experiment, a 2AFC task, the observers indicated which of two images with the same contrast variation but differing means had higher contrast. As the variation in contrast across the image increased, the apparent contrast of the image increased for both real scenes and phase-randomized versions of the same image. Additionally, sensitivity to change in the mean contrast of the image decreased when the variation in contrast across the image increased. The thresholds were higher for real scenes compared to phase randomized scenes even when their luminance distributions were equated. Equivalent noise analysis showed that the difference between sensitivity to real and random phase images was caused mostly by an elevation in local noise, with little systematic change in sampling efficiency across the four observers. These results suggest that local contrast responses are modulated by the presence of spatial structure and phase alignment across spatial scales in natural scenes which acts as an additional source of contrast variation.