Local spectral models for texture segmentation and estimation of textured surface inclination
Mark R. Turner, Ruzena Bajcsy · Scholarly Commons (University of Pennsylvania) · 1990
This dissertation focuses on two problems in visual perception: (1) segmentation of images containing differently textured regions, and (2) estimation of planar surface inclination by measurement of texture gradients. The models presented have a number of common methodological features. Each uses 2D Gabor filters having the same general form as that found to be a good description for simple cell receptive fields in visual cortex. Using this local spectral representation of image information, a higher order image analysis model is developed. In most cases, the algorithms allow parallel execution. Not only is this in accord with the psychophysical evidence for various "preattentive" processes in humans, it also permits rapid execution in machine vision applications on parallel architectures. Results of these investigations are presented in the following chapters. 1. Texture discrimination by Gabor functions. This paper suggests that textures preattentively discriminable by humans are, in most cases, easily segmented by simple statistical differences between combinations of Gabor filter amplitudes. 2. Iterative and cooperative models for segmentation. Several iterative and cooperative algorithms are described to segment images using Gabor filter amplitudes. 3. Neural assemblies as building blocks of cortical computation. This paper addresses the relationship between experimental observations of neural assemblies in cortex and models such as those described in the previous chapter which may identify computational functions of these assemblies. 4. Estimation of textured surface inclination by parallel local spectral analysis. A cooperative algorithm is presented for estimating the inclination of planar textured surfaces from the two-dimensional distributions of Gabor filter amplitudes. 5. Receptive fields for the determination of textured surface inclination. A neural network model, developed to solve the texture gradient problem, contains receptive fields which have an oriented structure in an unusual space/spatial frequency coordinate system. 6. Underestimation of textured surface inclination by human observers: A model. With monocular stationary views, human subjects in psychophysical experiments tend to underestimate the inclination of "irregular" textured surfaces. The cooperative algorithm exhibits this same behavior, suggesting that a similar computation may be occurring in visual cortex.