A neural network model for long-range contour diffusion by visual cortex

Abstract

We present a biologically plausible neural network model for long-range contour integration based on current knowledge about neural mechanisms with orientation selectivity in the primate visual cortex. The network simulates diffusive cooperation between cortical neurons in area V1. Recent neurophysiological evidence suggests that the main functional role of visual cortical neurons, which is the processing of orientation and contour in images and scenes, seems to be fulfilled by long-range interactions between orientation selective neurons [5]. These long-range interactions would explain how the visual system is able to link spatially separated contour segments, and to build up a coherent representation of contour across spatial separations via cooperation between neurons selective to the same orientation across collinear space. The network simulates long-range interactions between orientation selective cortical neurons via 9 partially connected layers: one input layer, four layers selecting image input in the orientation domain by simulating orientation selectivity in primate visual cortex V1 for horizontal, vertical, and oblique orientations, and four connected layers generating diffusioncooperation between like-oriented outputs from layers 2, 3, 4, and 5. The learning algorithm uses standard backpropagation, all processing stages after learning are strictly feed-forward. The network parameters provide an excellent fit for psychophysical data collected from human observers demonstrating effects of long-range facilitation for the detection of a target orientation when the target is collinear with another orientation. Long-range detection facilitation is predicted by the network’s diffusive behavior for spatial separations up to 2.5 degrees of visual angle between collinear orientations.