A NEURAL-NETWORK MODEL IN STEREOVISION MATCHING

The paper outlines a method for solving the stereovision matching problem through a Neural Network approach based on self-organizing technique. The goal is to classify pairs of features (edge segments) as true or false matches; giving rise to two classes. Thus, the corresponding parameter vector from two component density functions, representing both classes and drawn as Normal densities, are to be estimated by using an unsupervised learning method. A three layer neural network topology implements the mixture density function and Bayes's rule, all required computations are realized with the simple ''sum of product'' units commonly used in connectionist models. The unsupervised learning method leads to a learning rule, while all applicable constraints from stereovision field yield an activation rule. A training process receives the samples to learn, and a matching process classifies the pairs. The method is illustrated with two images from an indoor scene.