Computational Approaches to Die Analysis in Ancient Numismatics

Identifying the dies used to strike patterns on ancient coins and medals is crucial in numismatics and requires a thorough examination of design marks. This process is extremely challenging due to subtle differences among hand-engraved dies, which wear over time, complicating die identification. It is labor-intensive and demands meticulous attention to historical and minting details. In this thesis, we tackle die identification using agglomerative clustering to group coins based on similarities in surface features. We chose this method over deep learning approaches due to its reported superiority and the lack of large, annotated datasets required for effective deep learning training. To address this, we created a smaller dataset from the numismatic catalog Egyptian Hoards I: The Ptolemies by Julien Olivier and Thomas Faucher, focusing on 100 Ptolemaic tetradrachms and their obverse dies. We conducted an extensive performance evaluation to develop the optimal clustering scheme, explored various preprocessing methods, and assessed performance using a broader set of evaluation metrics not previously used in die analysis. Additionally, we proposed a new criterion for evaluating imbalanced datasets, which is particularly relevant for our case.