Reading Legends on Ancient Coins: An Object Detection Approach for Character Recognition on a Novel Roman Republican Dataset

This research addresses the challenge of automated character recognition on ancient Roman Republican coins, a task essential for historical classification and digital humanities. Unlike modern coins, these artifacts feature legends (inscriptions) that are highly non-uniform due to manual minting techniques, lack of standardized placement, and significant surface degradation from centuries of burial and corrosion. The author treats character recognition as an object detection problem, creating a novel dataset of 5,654 images with 38,808 manual annotations across 21 character labels. The study evaluates the performance of several YOLO (You Only Look Once) architectures—specifically versions 3, 4, 5, 7, and 8—to determine which can best handle the high intra-class variation and imaging inconsistencies inherent in archaeological artifacts.

The experimental results indicate that the YOLOv7 family of models outperforms other tested versions in this specific domain. YOLOv7-Large achieved the highest mAP50 score of 90.4%, followed closely by YOLOv7-Extended (90.2%) and YOLOv7-xl (90.1%). The study highlights that the non-uniformity of legend placement—which often varies horizontally, vertically, or along the coin border—poses a significant hurdle for traditional OCR methods, making the end-to-end object detection approach of YOLO variants particularly well-suited for this task.

Automated recognition of coin legends has profound implications for numismatics and archaeology. By enabling the systematic, large-scale analysis of coin inscriptions, researchers can more efficiently categorize artifacts, identify moneyers, and trace historical political propaganda. This work bridges the gap between deep learning and the digital humanities, providing a scalable framework that can help museums and historians manage vast, uncatalogued collections of ancient artifacts that would otherwise require labor-intensive manual inspection.