Purchase this article with an account.
Elizabeth Arsenault, Curtis Baker; The Role of Higher-Order Statistics in Naturalistic Texture Segmentation: Modelling Psychophysical Data. Journal of Vision 2010;10(7):1354. doi: 10.1167/10.7.1354.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Some texture boundaries are easier to segment than others, and characterizing these discrepancies is an important step in understanding the neural mechanisms of texture segmentation. Previously we demonstrated (Baker et al., VSS 2008) that contrast boundary segmentation thresholds in natural textures decrease when the higher-order statistics are removed by phase scrambling. We also demonstrated (Arsenault et al., VSS 2009) that naturalistic synthetic textures are subject to this phase-scrambling effect, and were able to determine that some higher-order statistics are more important than others. Here we sought to examine the extent to which a standard two-stage (filter-rectify-filter) model can account for the observed psychophysical data. Stimuli were naturalistic textures extracted from high-resolution monochrome photographs of natural scenes. Mean luminance and RMS contrast were fixed. A half-disc contrast modulation was applied to each texture to create a left- or right-oblique boundary. The first stage of the model consisted of a bank of Gabor spatial filters in a range of orientations and spatial frequencies. Each texture was convolved with this filter bank, subjected to a power-law nonlinearity, pooled, and passed through left- and right-oblique second-stage filters. The model simulated trial-by-trial results by making a 2AFC decision on the boundary orientation based on the second stage filter response magnitudes with additive noise. As in the psychophysical experiment, modulation-depth thresholds were obtained for two conditions: phase-scrambled and intact. The model is capable of producing results qualitatively similar to those measured in human observers: phase scrambling improves segmentation thresholds. This improvement can occur over a range of noise levels, power-law exponents (both compressive and expansive) and model architectures (both early and late pooling). These results suggest that a two-stage model, like human observers, can be sensitive to higher-order image statistics.
This PDF is available to Subscribers Only