Recognition of Handwritten Characters in Birch-Bark Manuscripts via Object Detection

The automatic transcription of medieval birch-bark manuscripts remains a bottleneck in Slavic paleography and digital humanities. This study adapts the YOLO-based object detection framework for detection and recognition of handwritten characters in digital photographs of Novgorod birch-bark fragments. A synthetic dataset of over 2,500 glyphs was created by extracting and denoising character fragments with a convolutional autoencoder, then compositing them onto varied-resolution canvases with controlled distortions. This corpus was used to train and benchmark SSD, Faster R-CNN, and YOLOv5–v11 detectors for 50 epochs. The best-performing model, YOLOv10, reached an initial mAP of 0.85 on synthetic data prior to hyperparameter optimization (HPO) and was further optimized—via anchor reconfiguration and learning-rate scheduling—to achieve mean average precision (mAP) 0.89. On a hold-out set of 1,100 real manuscript images, it attained mAP =0.79, with median Levenshtein, Jaccard, and Cosine similarity scores of 0.81, 0.73, and 0.79, respectively. An error taxonomy identifies common failure modes and informs model refinements. These results support the scalable transcription of birch-bark texts and enable quantitative studies of medieval Novgorod’s language, scribal practices, and document production. Future work will integrate Slavic-specific language models and multimodal imaging data to improve character recognition accuracy.