RJO and Peter Tompa wrote to let me know about a press release, Computer Recognizes Archaeological Material And Fake Van Goghs by NWO (Netherlands Organization for Scientific Research).
The press release discusses the PhD thesis of computer vision researcher Laurens van der Maaten. Via Google, I learned that last week van der Maaten defended his PhD thesis Feature Extraction from Visual Data in conjunction with a symposium on 'dimensionality reduction'.
His personal web site allows download of Matlab extensions that implement his techniques and a technical report on how to use it.
I wrote to Dr. van der Maaten who provided me with an paper, “A New System for the Classification of Modern and Historical Coins”. The historical coin portion of the paper discusses an experiment using 4822 Merovingen coins from the Dutch Money and Bank Museum (AKA The Netherlands Mint Museum or Het Nederlands Muntmuseum.) The researchers (van der Maaten, Paul Boon, and Eric Postma) trained their computer using the Merovingen coins and then fed 50 different Merovingen coins into the system. The goal was to find coins “perceptually similar” in the training database. The image above is taken from that paper (and modified by me to show a single row of pairs.)
So, rather than trying to identify the ruler or inscription, the task was to find coins that looked similar in the opinion of 10 non-numismatist humans. The paper reports that the computer got 20% right "magnitude", 43% right "orientation". I think orientation was figuring out how to orient the coin (right-side-up) but I am not certain.
The research was supported by NOW/CATCH, a Netherlands program to develop innovative tools for accessing cultural heritage.
In a personal email, Dr. van der Maaten reported that in a database of 4000 Roman coins the computer got a good match for 50% of the queries. I assume the criterion are similar, a match that looks perceptually similar to non-numismatists, such as a similar imperial portrait facing in the same direction. The Roman database used in the experiments is online, RICH: Reading Images for the Cultural Heritage. (I was previously unaware of this database and encourage Roman coin collectors to try it out and send me their thoughts in the collection, and its presentation and web interface.)
A success rate of 20% to 50% doesn't sound great but it is pretty impressive! As with the COINS results discussed on this blog, I am curious to see if statistical problems boosted the success rate. Most Roman coins feature a man's head facing right on their obverses. A system that classified EVERY image as a man facing right might be rated as 25% success by non-experts. I look forward to the publication of the Roman results in detail.
An image and 38 second video of the Roman coin recognition can be seen here. It is either a silent movie or my PC doesn't have the right version of Quicktime to play the sound.
Pagination of physical specimens and CSV downloads - I have made some minor modifications to the coin type pages in OCRE, CRRO, etc. A relatively small number of types across these corpora have more than 100 ...
4 hours ago