Automating coin analysis: AI for the University of Warsaw

Business context

The University of Warsaw conducts advanced numismatic research into Roman and Byzantine coins, where reliable die matching is essential. Previous manual methods were time-consuming, error-prone and difficult to scale as data volumes continued to grow.

The project involved analysing more than 7,000 coins attributed to nearly 50 rulers, with significant variation in material quality and representation. In collaboration with Leaware, the university chose to develop a scalable machine learning-based web application to increase the precision, throughput and repeatability of analyses, while enabling deeper interpretations of historical economic systems.

Challenge

The greatest challenge was a highly heterogeneous dataset: the coins varied in wear, lighting, photo background and iconographic detail. In addition, class distribution was uneven across rulers, making it more difficult to train conventional CNN models effectively and increasing the risk of overfitting.

The team had to build a solution capable of reliable die matching under conditions of noise and outliers, while reducing the impact of researchers’ subjective decisions. The solution needed to deliver repeatable results, browser-based usability, high performance and continuous improvement as new data became available, while maintaining the transparency standards required in academic research.

Solution

The result was a web application hosted on Azure with an Angular interface and a C# .NET backend, with machine learning components implemented in Python. The preprocessing pipeline includes background removal, greyscale conversion and cropping, among other steps, which normalise inputs and improve signal quality for the models.

Feature extraction was entrusted to a pre-trained CNN model, while die relationship identification was based on a hybrid of supervised and unsupervised methods. A two-stage clustering approach was applied, with dimensionality reduction using PCA followed by K-Means algorithms for cluster formation, DBSCAN for handling noise and outliers, and Gaussian Mixtures for capturing subtle patterns.

The architecture was designed for scalability and continuous improvement: the system learns from new data and automates repetitive tasks, reducing analysis time and increasing precision. Researchers gain a consistent working environment and can focus on interpreting results, setting a new standard for digital numismatic analysis.