A neural network solution for the correspondence problem

A neural network is presented to solve the motion correspondence problem. Prior to the neural network, a statistical correlation technique is presented briefly. This is a base for the neural network. In a preprocessing stage of image processing, features are extracted from two snapshots of a moving scene. Each feature can be described by a number of attribute values. To learn which features in both frames are truly matching, a network is set up. This network consists of units in five pools: one central pool and four attribute pools. Each central unit represents one possibly matching pair of features, and each attribute unit represents a fixed difference or ratio in attribute value of a pair of features. After updating the activations of all units several times, the interactive activation and competition network finds a solution for the motion correspondence problem