Recognition of Ancient Roman Coin with Alignment and Spatial Encoding

Roman coins play an important role to understand the Roman empire because they convey rich information about key historical events of the time. The ancient Roman coins were widely used to announce the achievements of Roman emperors to the public. They also served to spread messages of changing policies or merits through the empire. By engraving portraits on the coins, the Roman emperors also could show themselves to the entire empire. In short, the coins were the newspaper of the Roman empire. In this way, Roman coins are always connected to historical events and Roman imperial propaganda. Therefore, understanding the ancient Roman coins could serve as references to understand the Roman empire. Because the coin market is very active, many coins are traded every day, mostly over the Internet [1]. But ancient coins are also becoming subject to a very large illicit trade [5]. The traditional way to detect illegal traffic of ancient coins is to manually search catalogues, dealers or the Internet by the authorities. But the manual process has limitations and is too slow to cover all trade. Therefore, there is a need to develop both reliable and automatic methods to recognize the coins. In this paper, we address the problem of automatically recognizing who is on the ancient Roman coin, while leveraging their spatial structure and without specifically focusing on the understanding of textual transcripts on coins. The ancient Roman coins have regular structure: the location of the emperor is roughly at the center of the coin and the emperors share common aspects across different coins. However, coins of the same emperor also exhibit large variations. Some of these variations are due to the differences in the coin material and diverse state of coin degradation. Others are due to the differences in which the same emperor was depicted by different coin creators. These intra-class variability aspects make the task of recognizing the coins very challenging. To surmount these challenges we propose a framework to simultaneously leverage the consistencies in the coin structure and local appearance to improve the recognition accuracy. To this end, we investigate two approaches: a method based on discriminative deformable part models (DPM) [2], specifically adjusted to the coin domain through the use of polar coordinate representations and the Fisher vector [4] with spatial-appearance encoding. The model using DPM first detects the face of the emperor on the coin and uses the detected location to build a spatial pyramid. The Fisher vector based model directly encodes the spatial information in its representation. The use of both representations allows the recovery of consistent patterns that characterize different Roman emperors despite the outlined intra-class differences. We also introduce a large annotated database of Roman coins, consisting of over 2800 pieces made of different materials, depicting appearances of 15 Roman emperors [3]. This dataset allows us to establish the performance advantages of the proposed approaches compared with more traditional rigid spatial structure models such as the spatial pyramid.