Self-consistent training of a neural network with a step edge model for probabilistic edge labelling

Presents a robust neural network edge labelling strategy in which a network is trained with data from an imaging model of an ideal step edge. In addition to the Sobel operator, we employ preprocessing steps on image data to exploit the known invariances due to lighting variation 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 backpropagation algorithm is used in network training. A novel scheme for constructing the desired training set is proposed and results are shown for real images; comparison is made with the Canny edge detector. Several training sets of different sizes generated from the step edge model have been used to probe the network performance. Evaluation results for training and testing sets are shown