Detection of Orkhon-Yenisei runic inscriptions

The recognition of Orkhon-Yenisei runic inscriptions is an important yet highly challenging task due to the poor quality of source materials, the visual heterogeneity of runes, and the limited availability of annotated data. The deciphering process performed by archaeologists is prone to errors and subjectivity. The most difficult stage involves the segmentation and interpretation of symbols under conditions of artifact degradation and complex visual contexts. Objective. To develop and train a computer vision model for the automatic detection and classification of characters from the Kül Tegin runic alphabet on monument images, including both black-and-white copies and real color photographs. A prototype system was developed, combining two models: YOLOv11 for symbol detection and a convolutional neural network for classification. The detection model achieved mAP@0.5 = 0.825, recall = 0.801, and precision = 0.75. The classifier achieved 90.23% accuracy and an F1-score of 0.884. Feature visualization using t-SNE demonstrated clear clustering for most classes. Limitations were identified, including overfitting and class imbalance-especially when working with rare or unknown runes. The model can be adapted to other runic systems. The results can be applied in digital epigraphy and historical-linguistic research to enhance the speed and accuracy of ancient text analysis. The work lays the foundation for the development of more scalable and robust recognition systems for inscriptions under low-quality data conditions.