COGNITUS - Fast and Reliable Recognition of Handwritten Forms Based on Vector Quantisation

We report on an eecient intelligent character recognition tool for the automatic treatment of handwritten bank transfer forms. The classiication is based on nearest-neighbor algorithms and a novel binary clustering technique for the generation of large prototype sets. We introduce a new conndence measure which can be used on a decision tree structure to combine lowest error rates with a very high recognition speed. Likelihood vectors allow context correction by database queries based on dynamic programming techniques as well as an easy integration of diierent classiier approaches in a multi-agent environment. In this paper, we present all components of the prototype system and give details on its realization and on possible parallel implementations on embedded systems.