In this work we tackle the problem of automatic recognition of ancient coin types using a semisupervised learning method, namely Graph Transduction Games. Such problem is complex, mainly due to the low inter-class and large intra-class variations and the task becomes even more complex due to lack of labeled large datasets from certain ancient ages. In this paper we propose a new dataset which is chiefly the extension of a previous one both in terms of quantity and diversity. Moreover, we propose a game-theoretic model that exploits both sides of a coin to achieve higher classification accuracy. We experimentally demonstrate that proposed approach brings performance improvement in this complex task even when few number of labelled images are available.