Large-scale coin die studies with artificial intelligence – STUDIES

In the domain of numismatics and ancient economy, there is a long expressed and large expectation to estimate the volumes of produced coinage. Study of coin-finds and hoards not being reliable, die studies are the best way to understand the mint outputs and therefore the relation of coinage to Politics and History. The die study consists in comparing each coin of a series to define the number of engraved tools that were used to strike coins. For large series, it is humanly impossible.
In the project STUDIES, we propose to rely on recent advances in data harvesting, computer vision and deep learning to automatically estimate the similarity between coins, in order to simplify the die studies. Based on a full matrix of similarity, one can then identify clusters, each being a plausible candidate for a unique die. It will thus be possible to conduct die studies at a much larger scale, one to two orders of magnitude more than the largest studies to date.
Two sets of data will be investigated. The first one consists of die studies already implemented that will be used to train deep learning models. The second one is the data coming from both museums (curated) and auction sales (uncurated) that will be proceeded in order to create corpora of series of coins to be die studied. We will rely on research on data harvesting to be able to create corpora automatically whenever it takes now months for researchers to build it. Regarding the models, we plan to rely on both the image and textual data that is linked to the coins found on large collections. This work will further be used and developed to tackle online auction sales’ collections that provide millions of coin pictures and data.
Two tools will be developed during the project. The first tool will harvest institutional or online auction sites to collect large collections of coin images and associated textual descriptions and metadata. However, some of these sources (auction sites) have erroneous associated data. A scientific challenge in this context will be to correct these data automatically in order to align them properly with existing ontologies. The second tool will rely on a fine-grained classification approach to estimate the similarities between currencies automatically. We will then group the currencies to identify those corresponding to a common die. Two versions of the tool will be developed, the first one exploiting visual information alone, the second one benefiting from textual information and structured metadata.
These two tools will finally be integrated in a single software, with an interface and the appropriate documentation so that an expert in numismatics, not necessarily having advanced knowledge in programming, can use them easily. The whole will be made available under a free license.