Abstract
Analysis of local image statistics underlies a wide range of basic visual processes, including segmentation and surface characterization. Visual textures are useful probes of the neural computations that are involved, as they enable isolation of individual image statistics and detailed study of their interactions. However, while image statistics have enormous variety, most studies focus either on image statistics defined by multiple gray levels but ignore spatial correlations, or statistics that focus on spatial correlations but ignore gray levels. Recently, we proposed a model that goes beyond these limitations. The model is completely constrained by previous measurements: the impact functions of Silva & Chubb (2014) to account for sensitivity to multiple gray levels without spatial correlation, and the quadratic form of Victor & Conte (2015) that accounts for sensitivity to binary textures with spatial correlations. In an out-of-sample test, we (VSS 2016) tested the model for textures that combine 3 gray levels and spatial correlations. Its predictions were in reasonable agreement with perceptual measurements. Here, we further test the model with spatial correlations involving up to 11 gray levels. We examined two kinds of spatial correlations: "stepped gradients," in which the contrast of adjacent checks tended to increase gradually or decrease abruptly in one direction, and "streaks," in which adjacent checks tended to have the same intensity. Subjects (N=3) performed a 4-AFC segmentation task, in which target and background were defined by these features. For stepped gradients, thresholds were markedly higher for 5 gray levels than either for 3 or 11 gray levels. For streaks, thresholds showed little dependence on the number of gray levels and were lower overall than for stepped gradients. These findings were predicted by the model. However, there was a small anisotropy in sensitivity to vertical gradients, suggesting gradient-sensitive mechanisms that the model has not captured.
Meeting abstract presented at VSS 2018