Robust integration and detection of noisy contours in a probabilistic neural model

Contour integration is an important step in the process of image segmentation and gestalt perception. Experimental evidence with monkeys and humans demonstrates that this specific computation is performed very fast and highly efficient, even if contours are jittered, partially occluded, or reduced in luminance. Here, we investigate the reliability of a probabilistic algorithm for contour detection under various neuronal and environmental constraints, as e.g. synaptic noise or imperfect knowledge about the exact orientation of an edge at some position in the visual field. We show that under most conditions there exists a range of tuning widths for orientation-specific neurons in the visual cortex which yields an optimum in contour detection performance. In particular, we demonstrate an increase of the performance when the information of the orientation of the contour elements becomes more uncertain.