Extraction of salient contours in primary visual cortex: a neural network model based on physiological knowledge

A neural network for extraction of salient contours in visual images is presented. The network reproduces some typical characteristics of information processing in the primary visual cortex. Cells in the visual cortex are grouped to represent 100×100 distinct hypercolumns; each hypercolumn consists of 16 cells with different orientation preferences. Each cell in the cortex receives input from the lateral geniculate nucleus, arranged along the preferred orientation according to a Gabor function. Each cortical cell also receives a further feedforward input from inhibitory interneurons, and lateral connections (both excitatory and inhibitory) from the other cortical cells (feedback mechanism). Intracortical excitation is arranged according to experimental data, in order to implement the Gestalt proximity and good continuation criteria. Intracortical inhibition realizes a competitive mechanism among neural groups, to eliminate noise. Simulation results, performed on contrast visual images in the presence of large Gaussian random noise (standard deviation may exceeds 60% of contrast) demonstrate that the network can easily extract salient contours, by almost completely suppressing noise terms.