An experimental protocol for the performance evaluation of Optical Character Recognition (OCR) algorithms is described. The protocol is intended to serve as a model for using the University of Washington English Document Image Database-I to evaluate OCR systems. The plain text zones (without special symbols) in this database have over 2,300,000 characters. The performances of two UNIX-based OCR systems, namely Caere OCR v109a and Xerox ScanWorX v2.0, are measured. The results suggest that Caere OCR outperforms ScanWorX in terms of recognition accuracy; however, ScanWorX is more robust in the presence of image flaws.