Abstract
The analysis of texture patterns is at the heart of visual processing. ; Texture perception, texture segregation, and the perception of salient texture borders between perceptually coherent regions are traditionally linked to the notion of feature gradient (e.g., Nothdurft, 1985,1991; Landy & Bergen, 1991; Mussap & Levi ,1999). We recently argued (Ben-Shahar, 2006) that at least for the commonly studied Orientation-Defined Textures (ODTs) this link is fundamentally flawed since general ODTs of varying orientation generically exhibit salient perceptual singularities ; (i.e., perceptual boundaries between perceptually coherent regions) despite having no outstanding orientation contrasts. ; Although these singularities are extremely robust and consistent across observers, they defy ; not only popular texture segregation theories, but virtually all neural models (e.g., Li, 2000) and computational segmentation methods, either local ; (e.g., Malik & Perona, 1990) or global (e.g., Shi & Malik, 2000). Given this gap, in our previous ; theoretical account we argued that an appropriate ; approach to handle the segregation of ODTs should not (indeed, could not) rely on orientation gradients but rather on two orientation curvatures, one tangential and one normal, that come about naturally from considering the differential geometry ; of the ODT from an intrinsic point of view. Based on these ideas, here we present a novel biologically-plausible computational model that computes the two ODT curvatures from the output of V1 oriented receptive fields and continues to compute a perceptual singularity measure (Ben-Shahar, 2006) directly from any given ODT image. The results match human performance to great accuracy and provide an important first step towards a general curvature-based approach for the segregation of general textures.