System Design of an Integrated Pattern Recognition System or How to Get the Best Mileage Out of Your Used Pattern Classifier

摘要

The paper examines the conceptural design of a character recognition system. The overall, classification system is viewed as a network of three subsystems: a feature extractor, a classifier, and a contextual postprocessor. Some of the design dependencies that exist between these subsystems are always used in the construction of a pattern classifier. However, there are many instances where the subsystems are designed independently, ignoring the dependencies existing between them. When viewed in perspective, these dependencies are not as complex as might be expected and their consideration can result in substantial improvements in system performance. Integration of the network is achieved by two mechanisms: (1) introducing additional paths of data flow, and (2) modifying the design of each of the subsystems. This allows alternative methods for processing a pattern at different stages in the recognition process. The second means of integrating the pattern classifier is the main contribution of this paper. An attempt is made to redesign each of of the three subsystems on the basis of the design and performance of the other two subsystems. (Modified author abstract)