Abstract
Many early visual processes, including segmenting an image into its components and analyzing their surface properties, depend on local image statistics. Modeling the underlying visual computations is challenging, owing to the high dimensionality of image statistics and the potential for complex interactions. We therefore took an indirect approach, beginning with black-and-white textures with local spatial correlations. In this stimulus space, image statistics can be independently probed and fully analyzed, leading to a predictively accurate model of psychophysical sensitivities. Here, we show that key aspects of these findings extend to images with three luminance levels, and are also consistent with Chubb et al.'s (2007) studies of "scramble" textures with multiple gray levels. To investigate ternary textures, we developed a stimulus domain parameterized by the probabilities of all configurations of black, gray, and white checks in 2x2 neighborhoods (with gray halfway between black and white). This space has 66 free parameters. Each parameter corresponds to an image statistic: 2 first-order, 16 second-order, 32 third-order, and 16 fourth-order. Taking symmetry into account, the 66 statistics fall into 12 distinct categories. Sensitivity to these statistics was quantified in N=5 subjects, using a 4-AFC segmentation task. Stimuli consisted of 64x64 arrays of 14-min checks, containing a 16x64 target in one of four possible locations. We found that all first- and second-order statistics were salient, as were selected third- and fourth-order statistics. Consistent with findings for binary textures, results showed (i) nearly equal sensitivities to increments and decrements of an image statistic, and (ii) elliptical isodiscrimination contours indicating quadratic combination of signals. We also note that the Chubb model for discrimination of "scramble" textures with multiple gray levels has a probabilistic formulation that applies to textures with spatial correlation, and this formulation is at least qualitatively consistent with our key findings.
Meeting abstract presented at VSS 2017