Emergent segmentation of monocular visual invariants for space perception

A mechanism is proposed for the emergent segmentation of junctions with a priori unknown position, orientation, and order, i.e. number of inducing contrast edges, in static luminance fields. Emergent segmentation of visual scenes is a prerequisite for object recognition, obstacle avoidance, and other behavioral tasks of biological creatures with visual scenes. The authors describe results of an investigation into computational mechanisms allowing for real time response of a boundary oriented neural network architecture to straight and curved lines as well as to junction configurations of any orientation and any number of constituting contours. The proposed mechanisms are based on Grossberg's FACADE theory of preattentive segmentation and provide a functional extension of his boundary contour system. >