Methods for modeling the first layers of the retina

Abstract

Recently, analog preprocessing circuits modeling the first layers of the primate retina (low level vision) have been described. The author compares the operation of these circuits with traditional image preprocessing methods. Specifically, he investigates properties and limitations of a machine vision based edge detection method (inhomogeneous anisotropic diffusion) and makes a comparison with silicon retina architectures. Anisotropic diffusion is a modification of scale space filtering, disallowing smoothing across pronounced image gradients (which are supposed to reflect image edges). The silicon retina is an analog VLSI realization mimicking the function of the first layers of the primate retina. The silicon retina is shown to be faster and to provide better segmentation results than diffusion methods. >