How to compose a complex document recognition system

Abstract

The technical challenges in document analysis and recognition have been to solve the problems of uncertainty and variability. From our experiences in developing OCRs, business form readers, and postal address recognition engines, we would like to present design principles to cope with these problems of uncertainty and variability. When the targets of document recognition are complex and diversified, the recognition engine needs to solve many different kinds of pattern recognition problems, which are a reflection of uncertainty and variability. Inevitably, the engine becomes complex, raising a question of how to combine its subcomponents, which are not perfect in their accuracies. The design principles will be explained with examples in postal address recognition.