Simulation of an Optoelectronically Implemented Neural Network for Early Visual Processing

Abstract

Simulations of the operation of a neural network architecture for boundary segmentation, an important stage of early visual processing, are presented. The network is based on the CORT-X model, a multiple spatial-scale, feedforward architecture for boundary segmentation of noisy binary images. The CORT-X architecture can be made to accommodate, with slight modifications, existing optoelectronic hardware capabilities. Network performance is evaluated with respect to deviations from ideal response along two dimensions: (1) contrast ratio and (2) nonuniformity. The effect of these parameters on network performance as a function of the input image signal-to-noise ratio (SNR) is evaluated. With ideal response (perfect uniformity and infinite contrast) the modified CORT-X architecture performs as well as the original model. Finite contrast does not significantly degrade network performance as long as there is some reasonable contrast. On the other hand, nonuniformities as small as 10% degrade performance even for high SNR.

This work was sponsored by the Defense Advanced Research Projects Agency.