A network system for image segmentation

The authors describe a neural network for segmentation of a blurred and noise-corrupted image. There can be an arbitrary number of gray levels in the restored image. The simplest system found to do an acceptable job has several parallel networks detecting potential edges at different orientations in the image. Their output is combined in a final network, where the restored image is formed by filling in sections with appropriate gray-level values. To detect the edges, the parallel networks use directional second derivatives of the image, and they only differ with respect to which direction this derivative is taken. The authors find that at least two such orthogonal working networks are needed to do a reasonable segmentation. The system is tested on simple geometrical figures distorted by Gaussian blur and noise, and its performance is compared with that of other algorithms. The authors comment on the existence of similar structures in natural vision.<>