Church Slavonic Handwritten Text Recognition on the Example of the Ostromir Gospel (NRL, F.n.I.5)

Abstract

This paper discusses the development of a tool for Church Slavonic manuscript recognition based on the 11th-century Ostromir Gospel (National Library of Russia, F.п.I.5). The YOLOv8 neural network architecture was chosen for the task. The target objects for recognition are 42 Cyrillic letters and the simple titlo. We developed a heuristic algorithm that reconstructs text from an unordered set of objects detected by YOLO. The algorithm accumulates the text line-by-line from bottom to top based on the bounding box coordinates and outputs the result according to the Unicode standard. The final model was trained for 2 epochs on 1680 images containing mosaic compositions of characters taken from different pages of the manuscript. To reduce the imbalance in character instance counts, we inserted artificially created samples of rare letters such as fert. A Character Error Rate of 4.09% was achieved on the first 200 pages of the Ostromir Gospel (punctuation, spaces, and diacritics except for titlo were not included in the evaluation). The model learned to recognize most characters in the main text well, yet it struggles with small writing (e.g. in chapter titles) and similar letters and their combinations (including у when having a separate class for ѹ). Moreover, the text reconstruction algorithm proves unstable when the page is tilted and thus requires improvement. At this stage, we cannot claim that the tool is ready for a mass digitization of Slavic manuscripts, but the prototype shows promise for further development.