Machine Learning to detect fake coins

[JamesMason59]

Hi,

I was looking at some ebay coins and I had heard a lot of people saying to be cautious when buying from ebay.  I'm currently working with data analysis and use stuff like machine learning and I was wondering whether we could add another level of certainty when purchasing coins. Would it be possible to tell we have a fake coin from only a real picture. In the case of sovereigns, there are quite clearly some very bad fakes, but there are many which are incredibly difficult to tell apart. In the case of sovereigns, it's possible to spot a fake in some cases (specific years) by counting the number of edges in the rim for example which is incredibly difficult/tedious to do by hand/eye.

Therefore I was wondering whether anyone knows an approach to check this way using some software which can look at the color, size of the lettering/figures with the idea that a computer can judge better than a human eye can. Obviously the best checks would be physical such as weight, XRF etc. This software idea would only be to make a quick check and would not be foolproof. Do people think such an idea is feasible? A fanciful idea would be to make this into an app where you could simply take a photo of a coin and return something like 90% chance of being real.

Thanks

[PansPurse]

Presumably you'd have to do some kind of white /colour balancing on the photos too if you want to judge it based on colour (think there's a fair amount of colour variation in older sovereigns)

Really interesting idea though. I'm curious how much else you could train it on. A lot of the time dodgy eBay ads just sort of set off alarm bells with the way they are phrased or the quality of photos, the timing, seller feedback. I think you could probably add a lot of potentially helpful factors in

[JamesMason59]

Yeah, I'd imagine you would need to apply some sort of transformation on the images as people may take photos in different lighting,angles etc also but given enough data, there may be enough of a pattern/distribution in old sovereigns to tell whether a fake is an outlier (assuming we only look at color). Along different lines, the wear of the coin would need to be taken into account but I'm not sure how that would be accounted for. I wonder , is there a way to tell if a fake has been artificially worn?

[sovereignsteve]

This is probably do-able but you would need a huge database of genuine coin images to start with.

Fakes do sometimes have certain characteristics that crop up again and again but some go down a different route.

some show themselves when handled and measured and others defeat all but the most expert numismatist.

If this could be achieved it would be an incredible tool for use against the cancer of fakery but would take a lot of setting up and work to make it reliable enough for the wild.

I'm thinking this sort of thing would be best done in collaboration with an experienced dealer with a huge archive of photos. One such dealer springs to mind.

[JamesMason59]

@sovereignsteve, that's true. I'm not sure how familiar you are with what people currently do using convolutional neural networks (CNN) but usually you do need thousands of images (the more the better generally). This could be done by essentially taking images off google (provided you check sources etc.). However, a real problem could be getting a database of genuine fakes as there are thousands of examples of real sovereigns. If you could build up a good database of images for every sovereign by year and type, it could be possible to say whether a potential fake belongs to any of the categories and if it doesn't, it's a fake so I don't think the amount of data could be an issue. There will be a lot of variance though in the coins which is an interesting problem to solve.