Abstract
Line drawings contain most of the shape-related information in complex scenes. From a technical standpoint, however, automatic extraction of line drawings from raw images has proven extremely difficult, primarily because shape contours are influenced by widely dispersed visual features acting in subtle geometric combinations. To address this problem, we developed a recurrent network architecture inspired by the interconnection circuitry of primary visual cortex, which incorporates several representational biases tailored to the task of long-range visual contour integration. A learning scheme was used to train the modifiable parameters of the network to capture the statistical regularities of contour shape. Two separate inhibitory subsystems, one feedforward, one feedback, modulate contour predictions. We find that when the trained network is applied to complex images, well-organized contours are selectively boosted and texture edges are selectively suppressed, leading to a rough line-drawing-like sketches of the visual scene.