A neural network approach for combining processing of multiple early vision modules is described. Several energy functions are defined for computing intensity contours, optical flow, and stereo disparity. Hopfield neural networks are used for function minimization with continuation on the sigmoid gain function to avoid local minima. New vision integration approaches are developed by extending the work of Poggio and Gamble to include cooperative interactions between different vision modules and Hebbian learning of vision module coupling weights. Resulting algorithms facilitate fast, robust image segmentation and provide a nexus for recognition.
展开▼