From the University of Würzburg:
Historians and other Humanities’ scholars often have to deal with difficult research objects: centuries-old printed works that are difficult to decipher and often in an unsatisfactory state of conservation. Many of these documents have now been digitized – usually photographed or scanned – and are available online worldwide. For research purposes, this is already a step forward.
However, there is still a challenge to overcome: bringing the digitized old fonts into a modern form with text recognition software that is readable for non-specialists as well as for computers. Scientists at the Center for Philology and Digitality at Julius-Maximilians-Universität Würzburg (JMU) in Bavaria, Germany, have made a significant contribution to further development in this field.
With OCR4all, the JMU research team is making a new tool available to the scientific community. It converts digitized historical prints with an error rate of less than one percent into computer-readable texts. And it offers a graphical user interface that requires no IT expertise. With previous tools of this kind, user-friendliness was not always given as the users mostly had to work with programming commands.
The new OCR4all tool was developed under the direction of Christian Reul together with his computer science colleagues Professor Frank Puppe (Chair of Artificial Intelligence and Applied computer science) and Christoph Wick as well as Uwe Springmann (Digital Humanities expert) and numerous students and assistants.
Christian Reul explains the challenges involved in the development of OCR4all: Automatic text recognition (OCR = Optical Character Recognition) has been working very well for modern fonts for some time now. However, this has not yet been the case for historical fonts.
“One of the biggest problems was typography,” says Reul. One of the reasons for this is that the first printers of the 15th century did not use uniform fonts. “Their printing stamps were all carved by themselves, each printing house practically had its own letters.”
Error rates below one percent
Whether e or c, whether v or r – it is often not easy to distinguish in old prints, but software can learn to recognize such subtleties. To do so, it has to be trained on sample material. In his work, Reul has developed methods to make training more efficient. In a case study with six historical prints from the years 1476 to 1572, the average error rate in automatic text recognition was reduced from 3.9 to 1.7 percent.
Direct to OCR4all on GitHub