Towards automatic coin classification

Automatic image classification algorithms can support coin experts in their analysis and study of coins. These algorithms take digital images of coins as input and generate a class as output. Automatic classification proceeds in two stages. In the feature-extraction stage, the image is transformed into a compact representation that contains information on the presence of features. In the classification stage, the feature representations are mapped onto a class. This paper focuses on the first stage by presenting and evaluating two feature types for automatic classification of modern coins: contour features and texture features. For the second stage, a standard (nearest-neighbour or naive Bayes) classifier is used. We evaluate the classification performance obtained with both feature types on an image collection of modern coins. The classification results are promising. In addition, we test the performance on a collection of medieval coins. We show that the effectiveness of the features does not generalize to medieval coins, probably due to erroneous labelling of the images. The paper concludes by stating that automatic image classification algorithms may support coin experts in their analysis of modern coins. Future work is directed towards finding appropriate features for ancient coins.