A fast training approach to artificial neural networks designed for image segmentation

A novel training approach based on the backpropagation algorithm for image segmentation is presented. A set of training vectors is obtained by applying Karhunen-Loeve transformations on the training patterns. Training is started in the direction of the major components and then continues by including other components, in the order of their significance. With this approach, not only will the number of computations during training decrease, but also the problem of trapping in a local minimum will be minimized. This method is applied to image segmentation and compared to the general backpropagation algorithm