STEREOPSIS BY CONSTRAINT LEARNING FEEDFORWARD NEURAL NETWORKS

This paper presents a novel neural network (NN) approach to the problem of stereopsis. The correspondence problem (finding the correct matches between pixels of the epipolar lines of the stereo pair from amongst all the possible matches) is posed as a noniterative many-to-one mapping. Two multilayer feed-forward NN's are utilized to learn and code this nonlinear and complex mapping using the back-propagation learning rule and a training set. The first NN is a conventional fully connected net while the second one is a sparsely connected NN with a fixed number of hidden layer nodes. Three variations of the sparsely connected NN are considered. The important aspect of this technique is that none of the typical constraints such as uniqueness and continuity are explicitly imposed. All the applicable constraints are learned and internally coded by the NN's enabling them to be more flexible and more accurate than the existing methods. The approach is successfully tested on several random-dot stereograms. It is shown that the nets can generalize their learned mappings to cases outside their training sets and to noisy images. Advantages over the Marr-Poggio algorithm are discussed and it is shown that the NN's performances are superior.