FONT RECOGNITION BY A NEURAL NETWORK

Two neural network models, labelled Model H-H1 and Model H-H2 by Hogg and Huberman have been successfully applied to recognize 26 English capital letters, each with six font representations. These two models are very similar, but Model H-H2 has the capability for modification of the basins of attraction during the training phase, whereas Model H-H1 does not. This appears to be a desirable feature for a neural network. It is shown in this work that this is indeed true. In either model, it is difficult to find a single set of parameters for one network or memory that can distinguish all of the characters. Therefore, a cascade of memories is utilized. Thus, in the training phase, a decision tree is built by cascading the memory matrices that represent the models. That is successive layers of refinement in selection of basins of attraction are used to generate output patterns unique to each input pattern. In the recognition phase, the subject characters are recognized by searching in the tree. Model parameters such as memory array size, SminSmax, and MminMmax were varied to elucidate the model's behavior. It is shown that there exist parameter values for both models to achieve a 100% recognition rate when all six fonts are used both as the training and the recognition set. Model H-H2 significantly outperformed Model H-H1 in terms of recognition rate, use of memory space, and learning speed when all six fonts were used as the training set. © 1990 Academic Press Limited.