Study of multilayer perceptron accuracy improvement under fixed number of neuron

In this article multilayer perceptron accuracy improvement of real time recognition system under time limitations was explored. To solve this task two level tree of classifiers was used. The top level of the tree is a fast selector gotten without supervised training, and the bottom level is a set of neural net classifiers trained on the corresponding training sets. These scheme allows to increase the number of neurons used in recognition under the same processing time, that helps to increase generalisation power of the classifier. Recognition of embossed symbols on plastic cards was used as model task.