Adaptive step edge model for self-consistent training of neural network for probabilistic edge labelling

The authors present a robust neural network edge labelling strategy in which a network is trained with data from an imaging model of an ideal step edge. They employ the Sobel operator and other preprocessing steps on image data to exploit the known invariances due to lighting and rotation and so reduce the complexity of the mapping which the network has to learn. The composition of the training set to achieve labelling of the image lattice with Bayesian posterior probabilities is described. The back propagation algorithm is used in network training with a novel scheme for constructing the desired training set; results are shown for real images and comparisons are made with the Canny (1986) edge detector. The effects of adding zero-mean Gaussian image noise are also shown. Several training sets of different sizes generated from the step edge model have been used to probe the network generalisation ability and results for both training and testing sets are shown. To elucidate the roles of the Sobel operator and the network, a probabilistic Sobel labelling strategy has been derived; its results are inferior to those of the neural network