The automatic recognition and digitization of typewritten and handwritten texts requires sophisticated algorithms to produce usable results. The involved steps usually form a pipeline of several steps, i.e. the optical scanning of the text, segmentation of the letters, their classification and the context dependent disambiguation and correction of unknown characters. This task of finding a solution to assignment problems connects pattern recognition tasks with artificial intelligence research. A lot of commercial systems apply artificial neural networks. In multi-agent architectures that combine different pattern recognition techniques the performance of character recognition software is improved.
Research in our group aims for the improvement of the basic classification methods and their organization in a modular system. Our OCR system COGNITUS uses a hierarchical method of nearest-neighbor comparisons to a data base of 200000 typewritten and 60000 handwritten letters to perform the automatic scanning of credit transfer forms. Currently 10 forms per second can be analyzed. An associative data base is then used to disambiguate uncertain characters.
Application report (in German)")?>