Sunday, August 03, 2008

More on COINS

I've been thinking more about the EU's COINS project to find stolen coins (which I previously discussed.) I think it has a chance of success and will strongly incent collectors to keep better records.

There aren't a lot of details on the COINS site so I imagined myself as the EU's experts implementing a theft-detector. There are two key computer image algorithms that can be used. The first is an algorithm to tell if two images depict the same coin (or two mated cast fakes). This is a hard problem because of lighting and angle variations. Yet computers can look at things we would never look at, though, like tiny variations in edge geometry and centering. I predict a good chance of success for the Technische Universit├Ąt Wein group designing this algorithm or algorithms.

The second key algorithm is one to attribute a coin to a mint. This is much harder than the previous problem. I can think of three approaches that seem feasible (which I won't give today). One approach requires significant expert human time to train the algorithm. Numismatists used to be expensive, but now that Mumbai University in India is offering a 1-year Masters Degree in Numismatics for the equivalent of US$71 it shouldn't be hard to find experts.

Even without the second algorithm the COINS project could have a big effect. The second algorithm, which might have been applied to tell if a coin is from Iraq or Cyprus, is probably not needed in practice.

Imagine a COINS server spidering all commercial numismatic web sites to create a coins in commerce image database. The spider would save every image of every coin forever and associate it with the seller's ID. For some sites, like eBay, the buyer's ID could also be recorded.

When a coin is stolen police would upload pictures of the stolen coins into a registry (similar to VCoins' Stolen Coin Registry.) The duplication checking algorithm would run all stolen coins against all previous and future images in the commerce database. Running against older sales helps with cases like the Veliko Tarnovo museum heist, where it is alleged the stolen coins were sold on eBay before the theft was detected.

If the algorithm is fast it could be used by customs agents during routine checks. Agents would just wave a camera over a traveller's coin trays and if a match is detected it would beep. In this variation it's important that false positives be kept very low! ... but it is OK if not every match is caught.

Computers never sleep and might find stolen coins years or decades after humans have forgotten the theft. There are also old, unsolved thefts to find. I suspect collectors will start to only buy from sites and auction houses certified as “pre-spidered for buyer protection by the COINS system” — or from very trusted offline dealers. Even when buying person-to-person, many collectors will check against the COINS web site as part of due-diligence to ensure their new acquisition won't land their heirs in hot water.

It should also be possible for forgery fighters such as the ISBCC to load their database of fakes and casts instead of a database of stolen images upon their own servers with the open-source COINS software. The algorithm would then detect fakes and could alert authorities if the fakes aren't being sold with the word 'replica' in the online description. Forgery fighters could charge for the service of guaranteeing an online sale didn't match a known fake.

At first I thought a simple matching algorithm wouldn't help much against looting. I suspect was wrong... Currently when looting is reported the antiquities police send guards to the site. Guards are expensive, can be bribed, and sometimes fall asleep. A new approach is to deal with looters by taking a handful of mid-grade coins, entering them into the database as stolen, and leaving them at ancient sites! It's a lot cheaper to put $500 worth of coins on a site than guard it. No looter will be able to resist the salted coins, which will soon appear online or at a customs check point, be flagged by the comparison algorithm, triggering traditional police work to trace back to the looters.

No comments: