Low contrast object detection using a MLP network designed by node creation

In this paper we address the problem of detecting objects that are not clearly defined by an edge within the texture of an image. Multilayer perceptron networks using the backpropagation training algorithm are being used successfully as pattern classifiers for the object detection task. Although they have substantial benefits over conventional pattern classifiers, they do pose design problems and a widely used technique for obtaining an `ideal´ architecture is trial-and-error. In this paper we also propose a variant of the existing node creation methods, that uses a combination of a fixed number of iterations and cross validation as stopping criterion for one hidden layer networks