In this paper a hybrid handprinted character recognition system is presented. Four classifiers based on simple features work in parallel and their cooperation is used for quality improvement. The four classifiers are based on two different normalization sequences, on two different feature extraction methods and on two different classification techniques. The results of the classifiers are combined using three different methods. The first one is based on rules, the second method employes a weighed sum combination, based on a perceptron layer, while the third technique uses a multilayer perceptron acting as supervisor, which extracts the overall information contained in the output of the classifiers. The results obtained on the NIST Test Data 1 are reported using the upper-case letters in the NIST Special Database 3 as training set, featuring 4.20%, 3.74%, 3.68% error rate when no rejection is allowed, respectively for the three methods.
|Microelectronics Group Home Page||Staff|
|Research Activities||Teaching Activity|
|DEIS Home Page||University of Bologna Home Page|