Predictions from a computational model of texture perception compared with psychophysical data

Major theories of preattentive human texture perception due to Julesz [5,9] and to Beck [1,2] attribute preattentive texture discrimination to differences in first-order statistics of stimulus features such as orientation, size and brightness of constituent elements. These theories have typically been constructed for black and white dot or line patterns and are not directly applicable to grey scale images. An alternative approach [6,15, 4] has been to exploit the linear mechanisms (psychophysically observed spatial frequency channels and neurophysiologically observed blob, bar- and edge- sensitive neurons) which have been used to explain a range of phenomena in early spatial vision. Some experiments [4,3] suggest that this approach may explain texture perception better than the more symbolic, feature based approach of Beck and Julesz. However no scheme in this framework has been fully specified, implemented and successfully tested. The crucial experimental tests are the following: Does the model correctly predict the texture boundaries found preattentively by human observers, both in images of natural scenes and the synthetic stimuli from psychophysics literature? Even better, does it correctly predict the degree of discriminability for different texture pairs as measured by psychophysical experiments?