A computational model of texture segmentation

We present a computational model of human texture segmentation and argue for its utility in machine vision. Major theories due to Julesz and Beck attribute preattentive texture segmentation to differences in first-order statistics of stimulus features such as orientation, size and brightness of constituent elements. An alternative approach seeks to exploit psychophysically observed spatial frequency channels and neurophysiologically observed blob, bar and edge-sensitive mechanisms, and perform simple computations on the outputs of these to find texture boundaries. Previous models in this framework have been incompletely specified; our model is precisely stated and applicable to arbitrary grey scale textures. We claim that the responses of two types of mechanisms are necessary and sufficient: (a) center-surround mechanisms of various widths, and (b)oriented mechanisms of various widths and orientations which are even-symmetric about their axes. Simulation data on a number of texture pairs is presented