Handwritten text recognition system for processing census forms

This paper describes a document processing system for extracting and recognizing handwritten text entries on census forms. The system is intended to augment a manual key entry system, thereby lowering the total cost of transferring data from paper to computer-readable form. The system was tested on 9000 handwritten text fields taken from actual 1990 census forms. In this test, the system was able to achieve a field error rate below 4%, where error is defined as any single or multiple character deviation in system response from the actual response shown on the form. Most of the errors involved a small number of character deviations from the actual response, averaging less than one percent of the characters in the field. Anecdotal evidence suggests that these error rates are comparable to or less than those sustained during a manual keying process