Neural network classifier for character recognition

The purpose of this work is to develop a neural network classifier for the recognition of characters. The characters are presented to a digital computer though a CCD camera and a digitizer. The algorithm uses a gray level image of the character as input. The image is normalized. The feature used is stroke length. The stroke feature vector is calculated at each point and averaged over image subdivisions which are equal in area. This feature vector serves as the input to the neural network. A back propagation iterative learning procedure is implemented to teach single and multi-layered networks, i.e. to determine the network weights. A piecewise linear neuron characteristic was utilized as opposed to a sigmoid function. An estimation of the dependency of the learning procedure on the number of hidden nodes and the learning rate is made. The efficiency of the neural nets is investigated. The recognition rate obtained for the neural networks is compared to the rates obtained by standard nearest neighbour classifiers. All the algorithms were implemented in C on a MicroVax II workstation equipped with a high resolution image processing board.