If you Google “Republican Roman coins” you will be faced with thousands of fascinating images. There are three identifying features on most of them. Most show a head or object. Most have a mark identifying the mint. Most have a word known as the “legend.” The legend might be “ROMA” or “III VIR,” for example. This paper is about recognizing the legends found on Roman coins minted between 280 and 27 BC. These coins look old, and are often beaten up. Recognizing their legends defeats normal optical character recognition (OCR) techniques. This paper makes the case that reading legends on images of coins is almost as hard as decoding CAPTCHAs.

This paper should be required reading for researchers and grad students in scene text recognition (STR). It includes an in-depth survey of the field and adds new ideas. At first glance, I thought the proposed architecture (Figure 3) revisited Sedgwick’s Pandaemonium [1], but the solutions in this paper use state-of-the-art scale-invariant feature transforms (SIFTs) and support vector machines (SVMs). These are in turn guided by gathering features only in regions of interest (ROI). These are places in the image with large entropy. I would have preferred to see more detail about the features that were used, but this would make the paper too long. The resulting system achieves 25 to 65 percent recognition. We are still a long way from an app to help an amateur who stumbles across a coin in the back garden.