Feature Maps for Non-supervised Classification of Low-Uniform Patterns of Handwritten Letters

When input data is noisy and with a lack of uniformity, classification is a very difficult problem, because decision regions are hard to define in an optimal way. This is the case of recognition of old handwritten manuscript characters, where patterns of the same class may be very different from each other, and patterns of different classes may be similar in terms of Euclidian distances between their feature vectors. In this paper we present the results obtained when a non-supervised method is used to create feature maps of possible classes in handwriting letters. The prototypes generated in the map present a topological relationship; therefore similar prototypes are near each other. This organization helps to solve the problem of variance in the patterns, allowing a better classification when compared with other supervised classification method, a nearest-neighbor algorithm. The feature map was built using a Self-organized Feature Map (SOFM) neural network.