Discovering Similarities in Relief Patterns of Taranto Coins

Abstract

Die study, the comparison of coin reliefs to determine which coins were struck from the same die, is an important but time consuming task for numismatists. In this paper, we build a pipeline for automatic classification of normal maps of coin reliefs, consisting of a simple image registration algorithm and a Support Vector Machine classifier, and show ways to detect defects in coins given these classifications. Using the distribution of distances between edge pixels, we achieve a classification accuracy of 99.4% for a subset of coins, and 85.5% for the entire dataset. For defects, we show that we are able to detect cracks in the coins by comparing the gradients of pixels to neighboring pixels. We model a hammerstrike defect with a secondorder radial expansion and find that the effects of expansion are minimal, on the order of micrometers for a 12mmradius coin.