Texture-Based Methods for Analyzing Elementary Visual Substances

Human vision is generally assumed to embody a limited number of low-level, spatially parallel image transformations. These transformations define the basic properties vision can sense. Because texture discrimination tasks require rapid, spatially parallel processing, they are ideally suited to investigate these basic transformations. Here it is shown how to use paired comparison texture tasks to functionally analyze such transformations. Tasks are considered in which the observer is presented with two patches of texture, A and B, tiled with micropatterns drawn from a set Omega, and is asked to judge which is greater, M(A) or M(B), for some specific texture property M (e.g., M(A) might be the intensity variance of patch A). Performance is modeled by supposing that the observer approximates M(A) and M(B) by noisy, subjective quantities,;M(A) and;M(B), synthesized from the basic transformations resident in human vision. Psychophysical methods are provided for determining the differential impact on;M of the various micropatterns in Omega. Specifically, an occurrence in patch A (similarly for patch B) of a given micropattern omegainOmega is assumed to contribute an independent, additive random variable X(omega) to;M(A). The mean m(omega) of X(omega) gives the average impact on;M(A) of an occurrence of omega. It is assumed that the observer judges M(A)>M(B) on a given trial if;M(A)-;M(B)+Y>0, for Y a normal random variable with mean zero and unspecified variance. Simple, efficient techniques are provided for accurately estimating the shape of the function m: Omega-->ℝ, which defines the impacts of different micropatterns on M-judgments. These methods yield results that are invariant with respect to all unmeasured model parameters. MATLAB code is supplied for data analysis. Copyright 1999 Academic Press.