Approaches to the optimisation of pyramidal architectures for handwritten character recognition

This paper discusses the development of hierarchical image classifiers based on a pyramidal processing structure for application to the recognition of handprinted characters. Such classifiers have been found to be particularly effective in this type of application. Specifically, the paper investigates three ways in which the system designer may control the performance of the classifier: the introduction of a rejection mechanism to control the trade-off between overall classification rate and error-rate; the improvement of error-rate performance through the introduction of a mechanism for multi-source acquisition of classification information; and the use of a means of terminating the decision-making process at an early stage to improve classification throughput. Mechanisms such as these are important in providing the system designer with an appropriate toolkit of options with which to seek an optimal solution to a variety of practically-specified tasks