Image Based Recognition of Ancient Coins

Image Based Recognition of Ancient Coins is the title of a 2007 paper by Maia Zaharieva, Martin Kampel, and Sebastian Zambanini printed in Lecture Notes in Computer Science: Computer Analysis of Images and Patterns. Eight pages. The paper may be downloaded for $25 from SpringerLink.

The authors attempted to use computer programs to recognize Roman Imperial coin images. A database of 3000 Roman Imperial coin images was used (from the International Numismatic Commission; I don't think the database is online).

The authors attempt to use the techniques to automatically recognize modern coins (country of issue, denomination) from computer images that performed well at the “MUSCLE CIS Coin Competition 2006” (ERCIM News reports on that competition here).

The coin image database was manually divided into 106 'classes'. The authors don't say what a 'class' is; I'm guessing the class was the name of the emperor or empress. Classification software quality is usually measured by first setting it up using training database with information (the 'class', such as emperor name) to teach the system what to look for. Then a second test database of similar images is then fed into the software which must guess the 'class'. The guesses are checked against the classification done by a human expert. (The authors don't come out and say that is what they tried to do for ancient coins but I assume it was. That's how the MUSCLE competition was done.)

The paper's Abstract makes it sound like the authors hope to use computer analysis of coin images to fight commerce in the illegal antiquities market. They don't mention that automatic classification (to emperor, reverse type, etc) would be useful for numismatists in general!

The paper presents details on the mathematical techniques they tried.

The author's conclusion is negative: “features that achieve good classification ratio with modern coins easily fail with the classification of ancient coins.” (In other words, their software couldn't figure out which emperor was on their test database, if my reading of the paper is correct.) They were getting 4% success rates. Since they had about 100 classes chance would be about 1% so perhaps they were 4x better than chance.

The authors claim preliminary tests using a computer vision technique, Scale-invariant feature transform [wikipedia], shows promising results for medieval coins and they intend to attempt it against the medieval collection of the Fitzwilliam.

The authors have a follow-up paper, From Manual to Automated Optical Recognition of Ancient Coins, which is another $25. I haven't read it; if anyone buys it a summary would be very much appreciated!