EigenCoin: sassanid coins classification based on Bhattacharyya distance

A team of researchers has introduced EigenCoin, a novel computer vision framework designed to solve a specific, challenging problem in historical analysis: classifying ancient Sassanid coins from imbalanced datasets. The model, developed by Rahele Allahverdi, Mohammad Mahdi Dehshibi, Azam Bastanfard, and Daryoosh Akbarzadeh, constructs a specialized manifold and uses Bhattacharyya distance—a statistical measure of similarity between probability distributions—to map and classify coin images. This approach directly tackles the common issue in archaeology where examples of certain coin types are far rarer than others, which typically cripples standard machine learning models.

In their experiments, detailed in a paper submitted to arXiv, EigenCoin demonstrated significant performance gains. It outperformed other observed classification algorithms, achieving accuracy improvements ranging from 9.45% up to 21.75%. Crucially, the framework also showed a strong capability to handle over-fitting, a major problem where a model learns the noise in a small training dataset rather than the underlying pattern, making it useless for new, unseen data. By comparing holistic (whole-image) and feature-based approaches, the research provides valuable insights for applying AI to real-world cultural heritage projects where data is scarce and uneven.