I've written before about The COINS Project (Combat On-Line Illegal Numismatic Sales), a European Union sponsored effort “to enable the traceability of cultural heritage objects”. The project is developing “technologies aimed at allowing permanent identification and traceability of coins”.
COINS recently released software that attempts to compare and match coin photos against databases of coin photos. The newly released software comes in several pieces.
The key pieces are the Image Recognition Tool, said to be able to match a photograph against a database of coins, and the Shape Matching Package, which is apparently also for matching photos.
There is also a a web spider “to discover and identify stolen coins illegally sold on the internet” based on the open-source search engine Nutch. The spider comes with source code. It should robotically download all image files on a site and compare them against the database of coins. (My first reading of the search code did not impress me. I found code that looks for pages containing the word 'coin' or 'moneta' but I didn't find code to invoke image processing.) There is a user interface, which can be previewed here, for conducting simple searches of a coin database. For example, enter 'Alexander' in the simple query box, or 'eagle' in the iconographic query box. This Flash-based web interface is nothing special, and the spider's ability to search for stolen coins is only as effective as the photo-matching capabilities powering it.
A good image recognizer for ancient coins would be quite useful. It probably wouldn't help much to find particular categories of cultural property, such as ancient Cypriot and Chinese coins, as these are usually sold with descriptive text making photo recognition superfluous. Recognition technology would be useful to detectives for particular stolen coins and for computer-assisted die studies by numismatists.
I downloaded COINS' Image Recognition Tool (IRT) and loaded it with a database of 80 fake coins from my own photo-file. Although the IRT can compare images in popular formats such as JPEG the database only accepts greyscale TIFF files. I used Pierre Gougelet's free XnView to convert color JPEG files to greyscale TIFF.
I queried IRT using a fake image that wasn't in the database. I chose a cast fake Himera bronze, a type which gets past many dealers despite having been published as false in Wayne Sayles' Classical Deception. I had several of fakes from the same forger's mold in my database. The software coughed up five possible matches, with the highest being another coin of this type! (The other offered matches made little sense and were ranked poorly by the software.)
That seems like impressive performance. However, my database is small. Only 43 coins, or about half of the 80 coin database, correctly segmented as round. The other half of samples weren't exacted from the patterned background. Backgrounds that a human wouldn't even see, such as a finger, are not detected as expected by IRT.
The 4 non-zero matches returned in my first try represent 10% of the 43 coins whose edges were detected. Obviously a matcher that thinks 10% of the database matches a particular specimen is not yet trustworthy enough to assist law enforcement. Furthermore, the coin I first checked is a cast fake, so the geometry of the edge should help the software to detect matches more easily.
Oddly, the edge geometry didn't play a role. Although the correct match was found, the system believed a 60 degree rotation was needed. Actually the coins had the same orientation! (The picture at the top of this post is of this match.)
I tried genuine coins of different types than the fakes and got many false positives. The strength of the match is reported by IRT in the 'Matches' field. The two identical casts got 17 matches; a Ptolemaic bronze got 8 matches a shipping envelop with an eBay logo. I don't know what number is considered by the COINS scientists to be a good match.
I've just started playing with this software. I will try the Shape Matching Package next and report here on its accuracy. I invite other collectors experimenting with this software leave a comment on this post describing your success or failure matching coins using COINS' tools.
4 comments:
Very interesting report Ed. I personally collect a series of Roman Provincial coins by die variety. Even though I know precisely what I am looking at, I often find it extremely hard to distinguish die matches from die similarities. I have been very skeptical that any computer recognition system could do this as well as a human who can hold the coin in hand and change lighting characteristics and benefit from the extremely high resolution and color differentiation of the human eye. Yes, an Athens tetradrachm may be distinguished from an Alexander tetradrachm by image analysis, but identifying a specific coin with any certainty is going to be difficult if not impossible unless the comparison photos are identical -- which they will rarely be.
This is very useful to those of us working on the Project. The software passed various experimental tests, obviously, or it wouldn't have been released: details of those tests are in the reports downloadable at the Project's web-pages. But it's people like you and archaeologists in the field or in non-university contexts who will have to apply this stuff if it's to do the sort of good you envisage, so I'll definitely be directing the team members to this post.
Hi, I have downloaded the package years ago, unfortunately the website is now offline. Do you have any chance to upload the software? Regards
Sadly I don't have it. The COINS project's web site is gone, replaced by a squatter site. I was unable to find a new site hosting the software.
A few pieces of the old site can be visited using Archive.org's Wayback Machine. It had not occurred to me that a heavily-funded research project would not keep it's domain going or migrate.
I found a Github project, with some COINS papers and some coin recognition code at https://github.com/rjw57/ancient-coins. Perhaps contact the owner of that site?
Post a Comment