Local gain control and focal accommodation in the Self Similar Stack vision model.

Abstract

The aim of this paper is to report an extension to the computationally efficient Self Similar Stack model (Burton et al. Biol., Cybernet. 53, 397-403, 1986) to include the effects of local gain control in the retina. The method employed to do this has been to fit a family of difference-of-Gaussian functions to the human contrast sensitivity function curves of van Nes and Bouman (J. Opt. Soc. Am. 57, 401-406, 1967). The centre frequencies of the DoGs within each family are octave-related, in a simplified manifestation of the DoG channels found by Wilson et al. (Vision Res. 23, 873-882, 1983). The sensitivity of each level, or channel, that formed the original Stack model is modulated individually according to the fitted values, as the local illumination varies within an image. The model was tested against psychometric data obtained by Haig and Burton (Appl. Optics 26, 492-500, 1987) during experiments on visual discrimination. The consistency of the results indicates the validity of the approximations and the robustness of the model, either for machine vision purposes or for predicting human visual performance. A simple algorithm, developed for use with a machine vision application of this model, provides a means by which a TV camera may be focused automatically. The success of this algorithm, using the newly computed channel sensitivities, suggests that human focal accommodation may be regulated by a similar form of mechanism.