A neural-net computing algorithm for detecting edges in a gray scale image

The algorithm presented features simple connections between neurons and is highly effective in detecting edges in noisy gray-scale images. The algorithm consists of two steps: image filtering and edge detection. The image filtering portion detects the zero crossings of D/sub i//sup 2/G (x, y) * l(x, y) where D/sub i//sup 2/G (x, y) represents the second derivative taken in the ith direction of a 2-9 Gaussian function. The output is then normalized and used to initialize the ith layer of a multiple-layer, Hopfield-type, neural network with simple localized connections. During edge detection, processing elements either excite (cooperate with) or inhibit (complete with) processing elements in other positions and layers. The competitive-cooperative process among the processing elements builds up edges and eliminates noise. The algorithm was simulated by using only four orientations: 0 degrees , 45 degrees , 90 degrees , and 135 degrees . Simulation results show the algorithm's performance on images with varying levels of noise. The output of this algorithm could be suitable input for a pattern recognition network.<
>