Vom Bildnis zum Bit. Deep Learning zur Analyse der Ikonographie antiker Münzen

The classification of ancient coins remains a central but highly labor-intensive task in numismatics, as it relies on heterogeneous descriptive traditions and the manual comparison of large corpora. Although automated image recognition has been explored for several decades, practical applications have so far been limited by insufficient data standardization and the complexity of interpretative iconographic descriptions. This study investigates the potential and current limits of an AI-based approach to the automated classification of Roman coins. Focusing on Roman coinage (ca. 320 BCE–491 CE), the research combines Natural Language Processing and computer vision to address both textual and visual sources. Interpretative coin descriptions from digitized GLAM collections were transformed into machine-readable data using a custom Named Entity Recognition system based on a knowledge graph, incorporating alternative spellings, fuzzy matching, and targeted data augmentation. In parallel, coin images were preprocessed and used to train a Convolutional Neural Network to recognize the most frequent iconographic representations. The results demonstrate that reliable automated recognition of iconographic motifs is achievable for standardized coin images, while also revealing limitations caused by descriptive variance, image quality, and mild overfitting in the recognition of persons and attributes. The study shows that high-quality, interoperable data are a prerequisite for successful automation and that explainability and statistical robustness remain key requirements for practical use. Overall, the proposed approach provides a scalable foundation for automated coin classification in Roman numismatics and establishes a benchmark for future work. It also highlights perspectives for extending the method to other numismatic corpora, improving data interoperability, and supporting both academic research and heritage documentation through AI-assisted workflows.

Level thesis:degree_level

master

Grantor dc:publisher

Universität Heidelberg

Year

2025

Author dc:creator

• Celisi, Mattia

Contributors dc:contributor

• Arendes, Cord

Repository record source_url

http://www.ub.uni-heidelberg.de/archiv/37946

OAI identifier oai:identifier

oai:archiv.ub.uni-heidelberg.de:37946