Application of Super-Resolution Models in Optical Character Recognition of Czech Medieval Texts

BANKOVIČ, Mikuláš, Vít NOVOTNÝ and Petr SOJKA. Application of Super-Resolution Models in Optical Character Recognition of Czech Medieval Texts. In Horák, Rychlý, Rambousek. Recent Advances in Slavonic Natural Language Processing (RASLAN 2021). Brno: Tribun EU, 2021, p. 11-18. ISBN 978-80-263-1670-1.

Basic information

• Original name: Application of Super-Resolution Models in Optical Character Recognition of Czech Medieval Texts
• Authors: BANKOVIČ, Mikuláš (703 Slovakia, guarantor, belonging to the institution), Vít NOVOTNÝ (203 Czech Republic, belonging to the institution) and Petr SOJKA (203 Czech Republic, belonging to the institution).
• Edition: Brno, Recent Advances in Slavonic Natural Language Processing (RASLAN 2021), p. 11-18, 8 pp. 2021.
• Publisher: Tribun EU

Other information

• Original language: English
• Type of outcome: Proceedings paper
• Field of Study: 10200 1.2 Computer and information sciences
• Country of publisher: Czech Republic
• Confidentiality degree: is not subject to a state or trade secret
• Publication form: printed version "print"
• WWW: Domovská stránka workshopu Full text PDF
• RIV identification code: RIV/00216224:14330/21:00119900
• Organization unit: Faculty of Informatics
• ISBN: 978-80-263-1670-1
• ISSN: 2336-4289
• Keywords in English: Super-resolution; Optical character recognition; Medieval texts

Abstract

Optical character recognition (OCR) of scanned images is used in multiple applications in numerous domains and several frameworks and OCR algorithms are publicly available. However, some domains such as medieval texts suffer from low accuracy, mainly due to low resources and poor quality data. For such domains, preprocessing techniques help to increase the accuracy of OCR algorithms. In this paper, we experiment with two super-resolution models: Waifu2x and SRGAN. We use the models to reduce noise and increase the image resolution of scanned medieval texts. We evaluate the models on the AHISTO project dataset and compare them against several baselines. We show that our models produce improvements in OCR accuracy.