A modular hierarchical neural network for machine vision

An attempt is made to use a modular hierarchical architecture for image processing. A low-level feature detector (e.g., vertical edges) operates on a small patch of the image and generates outputs giving the probability that the feature is present and noting its position. An array of these detectors feeds a second-level network which computes feature probability and position over a wider field of view. The second-level units can be cascaded to third and fourth levels to determine global feature location and existence. This architecture can be implemented in VLSI electronics to provide real-time image processing over a large field of view