New Approaches to OCR for Early Printed Books

Abstract

Books printed before 1800 present major problems for OCR. One of the main obstacles is the lack of diversity of historical fonts in training data. The OCR-D project, consisting of book historians and computer scientists, aims to address this deficiency by focussing on three major issues. Our first target was to create a tool that identifies font groups automatically in images of historical documents. We concentrated on Gothic font groups that were commonly used in German texts printed in the 15th and 16th century: the well-known Fraktur and the lesser known Bastarda, Rotunda, Textura und Schwabacher. The tool was trained with 35,000 images and reaches an accuracy level of 98%. It can not only differentiate between the above-mentioned font groups but also Hebrew, Greek, Antiqua and Italic. It can also identify woodcut images and irrelevant data (book covers, empty pages, etc.). In a second step, we created an online training infrastructure (okralact), which allows for the use of various open source OCR engines such as Tesseract, OCRopus, Kraken and Calamari. At the same time, it facilitates training for specific models of font groups. The high accuracy of the recognition tool paves the way for the unprecedented opportunity to differentiate between the fonts used by individual printers. With more training data and further adjustments, the tool could help to fill a major gap in historical research.