Artificial neural networks for boundary extraction

Abstract

Artificial neural networks are designed to detect edges and extract boundaries. The system can accomplish the following tasks: 1) obtain enhanced boundaries; 2) recover missing edges; and 3) eliminate false edges caused by noise. The research comprises two phases, namely, boundary extraction by a BP net and boundary enhancement by a modified Hopfield neural network. The BP net is trained by 560 typical boundary patterns to enable the network to determine the boundary elements with 8 orientations and to provide the boundary measurement for further processing. A modified Hopfield net is proposed to enhance boundary measurement. Based on constraint satisfaction and the competitive mechanism, interconnection between neural cells are determined. A criteria is provided to find the final stable result which contains the enhanced boundary measurement. The neural network was simulated on a SUN Sparc station. Test images were degraded by random noise up to 30% of the original images. Comparing with the Gaussian edge detection and optimum edge detection, the results are very promising: boundaries were extracted, noise was eliminated, and boundary elements missed in other methods were detected.