Training edge detecting fuzzy neural networks with model-based examples

A model-based method for training feedforward, backpropagation neural-like networks to produce edge images from data such as forward looking infrared and gray tone pictures is presented. The authors´ approach is to train the network on a very small basis set of binary-valued window vectors which are first scored using the Sobel edge operator. Sobel scores are then used to select training vectors that have either crisp or fuzzy edge labels. This training scheme is independent of all real images. The method proposed is illustrated by comparing FF/BP edge images with those produced by the Sobel and Canny edge operators