Abstract
As pointed out by Purpura et al. [Proc Natl Acad Sci USA 91, 8482 (1994)] extracting information about edge and object relationships is equivalent to computing higher order correlations between image regions. A useful tool for studying our sensitivity to such correlations are the isotrigon textures. While groups of these textures are readily discriminable from each other collections of these textures have identical average first to third order spatial correlation functions. We show that the 18 isotrigon texture types considered have characteristic populations of mini-textures: small subdomains of the parent textures. We also determined human discrimination performance for 27 types of isotrigon texture pairings. These discriminations were determined for 4 pattern sizes, 2 check sizes and 2 contrasts, in up to 5 subjects. Discrimination performance was compared with outputs from statistical discriminant models. The input to the discriminant models was 5 variance measures derived from the mini-textures. For the discriminant analyses mini-textures were pre-multiplied by oriented or unoriented receptive field models and variance measures were computed on these outputs. The variance measures included 2D oriented and unoriented versions of the Allan Variance, and these are shown to be related to physiologically plausible mechanisms for calculating forth-order spatial correlations. Our quadratic discriminant models behaved similarly to humans suggesting that humans discriminate these textures as if they were using quadratic comparisons between the variance of the output of oriented filters, where the variance is either computed either over quite large parts of the image or more locally with an orientation bias.