Improving edge measurement on noisy images by hierarchical neural networks

Abstract

To improve edge operator measurement, a hierarchical neural network is proposed in this paper. The neural network is designed to adjust the edge measurement based on the neighboring edges. The overall strategy is to analyze the local edge patterns to reveal and reinforce the curvilinear edge structure while suppressing unwanted random noise. The hierarchical network consists of three levels. At the first level, eight selective functional-link nets associated with eight edge directions are working in parallel. Each of the nets determines the potential adjustment on the edge of concern according to the selected processes and the input local edge pattern. The second level is a Maximum Detection Subnet which makes a decision to alter the gradient magnitude and the direction of the edge element of concern in agreement with the local surrounding. The last level computes the adjustment for the edge measurement. Using the neural network for each edge element, highly parallel processing can be achieved. To apply global analysis, an iterative approach is incorporated with the neural network. As a result, the final edge measurement is more accurate.