Abstract

A growing trend in machine vision is the convergence of biological and computer vision approaches to understanding the basic organization and principles of image-analyzing systems. Insights into how the eye and brain organize visual data may provide novel and powerful computational paradigms for image processing. New evidence is presented that retinal receptive fields in primates and cats are functionally equivalent to filters defined as a weighted difference of a Gaussian function and its Laplacian. A new and extremely fast algorithm for approximating such filters via a difference-of-offset-Gaussian (DOOG) mechanism is presented. The technique is faster than the usual difference-of-Gaussian (DOG) mechanism using two concentric Gaussians, since only a single Gaussian convolution is required. Third- and fourth-order oriented Gaussian derivativelike filters, which are commonly seen in biological systems at visual cortical levels, can also be easily formed by additional lateral difference processing. Gaussian derivatives are known to be near-optimal filters for the simultaneous detection and localization of intensity discontinuities such as lines and edges in the presence of noise. The features and benefits of a machine vision system based on multiple-order Gaussian derivative filters are discussed.

© 1985 Optical Society of America