Segmenting textures using cells with adaptive receptive fields.

Abstract

Textural segmentation plays an important role in the figure-ground discrimination process. Evidence from neuroscience and psychophysics suggests that the segregation of texture patterns composed of oriented line segments is strongly influenced by the orientation contrast between the patterns (Nothdurft, Vision Res. 31, 1073-1078, 1991). In contrast to models available in the literature, this paper presents a neural network architecture for textural segmentation that can adaptively delimit the boundaries of uniformly textured regions. Cells with adaptive receptive fields encode uniformly textured regions by diffusively interpolating the estimates of feature orientations across the image. Orientation contrast boundaries are detected at gradients across the interpolated regions to produce the final segmentation. Consistent with neurophysiological data from simple and complex cells sensitive to static and moving textural patterns (Hammond and MacKay, Exp. Brain Res. 30, 275-296, 1977), the present model suggests that preattentive textural segregation can be performed by early visual processes localized to areas V1, V2, and perhaps V3 or V5. The computational results further support the idea that textural segmentation should not be thought of as a 'static' process, but as a system that employs cells with context-dependent response characteristics (Gilbert and Wiesel, Vision Res. 30, 1689-1701, 1990), here modeled as adaptive receptive fields.