High-order statistically derived combinations of geometric features for handprinted character recognition

An intelligent character recognition (ICR) system for offline recognition of isolated handprinted characters is presented. The system comprises a feature extraction stage and a classifier (a feedforward network) that is trained using error backpropagation. The feature extraction stage consists of two steps. Given a bilevel image of a character, the system first extracts a set of experimentally optimized raw geometric features. These features are then combined to obtain higher order features which have higher discriminating ability according to the Fisher discriminant measure. The specific combinations of raw features that are chosen in particular application are determined by statistical analysis of the training data. The nonlinearity introduced by the feature combination is found to be much more powerful than the traditional sigmoid nonlinearity of backpropagation neural network classifiers. It is shown that by using the high-order features, one can drastically reduce the number of hidden nodes required in the classifier while retraining the same level of classification accuracy