Key points are not available for this paper at this time.
The recognition of hand-printed numerals is studied on a broad experimental basis within the constraints imposed by a raster scanner generating binary video patterns, a mixed measurement set, and a statistical decision function. A computer-controlled scanner is used to acquire the characters, to adjust the raster resolution and registration, and to monitor the black-white threshold of the quantizer. The dimensionality of the decision problem is reduced by a hybrid system of measurements. In the measurement design, three types of measurements are generated: a set of "topological" measurements, a set of logical "n-tuples," both designed by hand, and a large set of n-tuples machine generated at random under special constraints. The final set of 100 measurements is selected automatically by a programmed algorithm that attempts to minimize the maximum expected error rate between every character pair. Computer simulation experiments show the effectiveness of the selection procedure, the contribution of the different types of measurements, the effect of the number of measurements selected on recognition, and the desirability of size and shear normalization. The final system is tested on four data sets printed under different degrees of control on the writers. Each data set consists of approximately 10 000 characters. For this comparison, a first-order maximum likelihood function with weights quantized to 100 levels is used. Error versus reject curves are given on several combinations of training and test sets.
Bakis et al. (Mon,) studied this question.