Optimum Design Parameters of Classifiers Used for Omni-font Machine-printed Numeral Recognition based on the Minimum Classification Error Criterion

摘要

The optimal design parameters of classifiers for omni-font machine-printed numeral recognition based on the minimum classification error (MCE) criterion are determined experimentally. The design parameters that influence the accuracy of an optical character reader (OCR) are: similarity measure (or distance measure), kinds of features, dimension of the feature vector, method of training, number of templates per category, and the size of a training sample set. It was found that the optimum design parameters were simple similarity, four templates per category, and 576 dimensions (i.e., four directional feature planes of 12 x 12 blocks). The directional feature classifier with these design parameters gave the best performance and had the smallest memory size and computational cost of all the classifiers.