Experiments with Kohonen´s learning vector quantization in handwritten character recognition systems

In this paper we present the results of classification of handwritten characters on a Kohonen neural network. Three types of features, Fourier transform, geometric moments and shadow feature extracted from handwritten character data were used for classification. Classification accuracy is found to be much higher with the shadow feature in comparison to the more traditional Fourier transform and geometric moments. We have also explored the relation between Kohonen´s learning of orientation based correlations and the learning rule of a minimum distance approach, used in a feedforward Athena neural network