Hierarchical artificial neural networks for edge enhancement

Abstract

Hierarchical artificial neural networks are designed to enhance edge measurement. The neural network comprises three subnets which are connected in a hierarchical manner. The edge Contour Detection subnet determines the orientation of the most probable edge contour (if one exists) in the local edge pattern, the Gradient Adjustment subnet makes a decision to alter the gradient magnitude, and the Orientation Determination subnet adjusts the direction of the edge element of concern to be in agreement with the local surroundings. In order for the neural network system to perform correctly and accurately, each of these subnets must acquire suitable weights by learning. In our learning algorithm, a modified generalized delta rule incorporates a momentum term to achieve fast learning without oscillation. The neural network is simulated on a MIPS M120-S machine running UNIX. Test images are degraded by random noise up to 30% of the original images. The results are very promising: true edges are detected and enhanced, false edges are suppressed, noise is eliminated, and missing edge elements are interpolated.