Line Detection in Nomenclator-Encrypted Historical Manuscripts

In this work, we present an approach for text line detection in historical manuscripts encrypted with a nomenclator cipher using transfer learning with the YOLOv8 architecture. We compare the performance of all official model variants on the ICDAR cBAD 2019 dataset and then evaluate various fine-tuning strategies on a proprietary in-house dataset of nomenclator ciphers. The best result was achieved by the YOLOv8x model with fine-tuning applied to both the head and neck, reaching an F1 score of 0.9423. The results confirm that partial fine-tuning significantly improves accuracy without requiring full retraining.