| Author |
 Topic  |
OldSkoolMadSkilz
Pillar Of The Community
United States
777 Posts
|
Yesterday I got to thinking that there must be a way to stop counterfeiting. My idea was to compare an image of a coin to a known authentic image. Look at the attached screenshot for what I have so far.
The two images at the left are the images being compared. The next image over is each of the images to the left cycled back and forth. The slider below controls the speed of the cycle, from twice a second to 20 times a second. It showed some differences that I would have missed otherwise. The more worn coin is much more deeply struck. I think all but the most trained graders would miss this.
The next column of images is a digital comparison. The top is the first minus the second, below, the second minus the first, and below, it's a secret. In these, the more similar the images, the more gray they appear. Notice how the LIBERTY and date nearly disappear in the bottom one. The hair detail shows because it is very different between the two because of condition.
The last column is an illustration of edge detection. I was a bit disappointed in the results. One might think that the top image shows few edges because of wear. Not so. The algorithm I used to rotate the coin altered the image data in such a way as to confuse the edge detection algorithm.
This procedure requires each coin to be in nearly the same exact position. I'm working on a way to do that. Also, it's very sensitive to the lighting on the coin. Not much I can do about that.
So, would this software be beneficial to the numismatic community? Is there already something out there?
|
|
|
|
Pillar Of The Community
United States
2876 Posts |
One thing to keep in mind, not all coins of the same date look the same. Only coins coming from the same die pair are. You could end up with a dozen different coins of the same date depending on how many dies were used. For example to detect a counterfeit large cent from an original large cent of the same date, you would need to check it comparing "like" Sheldon or Newcomb numbers and for Morgan and Peace dollars by the VAM's. Might turn into a big job to get this to work. And if you do, getting the TPG's on board first would go a long way in being able to market it commercially. |
|
Pillar Of The Community
United States
777 Posts |
Yes, I've considered that. I was thinking of making one "coin A" image and multiple "coin B" images to allow for multiple comparisons. What you see is about 4 hours work, but the auto align code will be a bear. I may go the TPG route because the software needs a large library of known good coins to work. Might be something to be web enabled. Edited by OldSkoolMadSkilz
06/12/2012 2:54 pm
|
|
Pillar Of The Community
United States
2876 Posts |
You definitely have your work cut out for you. Good luck with your project.
|
|
Moderator
6464 Posts |
I was at a coin show where they were selling a machine that took pictures of your coin and scanned it to enable you to tellit from any other coin. This was for the purposes of identification. Cost of the machine was a little high for what I had in mind and it did not differentiate original from counterfeit coins, but the industry is getting better due to the new ideas of collectors each day
rggoodie
aka Richard
www.aussie-coins.com
"catch em doing something right"
|
|
Pillar Of The Community
United States
5336 Posts |
If the coin type software had "compare-to-pictures" of all available die marriage possibilities, it definitely could work! And! Not only would the software detect counterfeits, but it would also automatically ID the Newcomb, Sheldon, VAM, Overton, etc of the coin!! I'd invest in that!  |
|
Pillar Of The Community
United States
3294 Posts |
I would be very interested in such software once it had been proven effective.
|
|
Pillar Of The Community
United States
892 Posts |
Very interesting approach. I'm a computer scientist and programmer and have thought about this but was seeing lots of tech problems. However, those of us in technology are expected to solve those problems and so I admire your enthusiasm.  Having the images of all possible varieties is a start. However, I wonder about the condition of the coin you are trying to validate. If a key component of identifying the coin is worn (and doesn't match) - how will the algorithm distinguish that from a fake component? This looks like a huge recursive algorithm. Please keep us in the loop on your progress. |
|
Pillar Of The Community
United States
777 Posts |
Cipster
Good! Someone to brainstorm with.
Yes, condition is an issue. For digital identification, the points where the relief and fields meet will provide significant matching. Or I should say, make a mismatch stand out. The details are important, but difficult to quantify digitally. That's where the strobing between the two images comes into play. The human eye and brain can process those details much better. Consider that the same coin lit from different angles will look quite different digitally, but much the same to the eye.
Actually, the digital process isn't that bad. I step through the image comparing pixels. The tough part will be finding a way to align the two images. On well photographed coins, I just need to manipulate x, y, and rotation. For poor photographs, I have to compensate for tilt and skew. I'm considering one option where the user places three reference points on the images and the software adjusts the images to make them match. I can tell when the optimum overlay happens by looking at the total match, but as far as using this to tweak it in place, it only applies when the match is already close.
|
|
Moderator
United States
14435 Posts |
I think the only problem with this is you would have to have every known variety listed because on some coins the design does look a little different in places and may show up and make you think its counterfeit when in fact its not, its just a different variety of the same type and maybe even the same date. But any tool a person can use that will educate them is a good thing though
|
|
Pillar Of The Community
United States
777 Posts |
The images are not stored in the software, You would paste each into the comparison windows, so you'd pick the images from the web or whereever. I'm thinking of putting multiple "coin B" slots in the UI.
I highly evolved version of this software would be server based. You'd snap a photo with your smartphone, send it to the server and it would report back whether it was authentic and what it matches. That's a few years out.
|
|
Pillar Of The Community
United States
1262 Posts |
This is a magnificent idea that I've toyed around with a bit myself using scripted GIMP plugins and node.js.
However, after a lot of effort, the biggest problems I came across were with lighting, toning, keystone/orientation, and wear.
Out of all of those problems, keystone and orientation is the easiest thing to fix as there are orientation and frame detection algorithms out there that work pretty well. The most fun I've messed with is that I trained up a neural network to recognize not just the orientation of a coin, but the type of coin (just with modern Pennies, Nickels, Dimes, and Quarters [obverse] -- state quarter reverse recognition would have to be its own network, ugh..).
Outside of that, straight image comparison math is very limited. What you'd really need to do is take a heightmap of the actual coin (with a high-def stereo camera or laser scanner) and then you can do a height subtraction with noise reduction, and perhaps develop some sort of wear transformation model to narrow down what areas to focus your comparisons on.
In the end it'd need to be more like a fingerprint recognition program where known points of interest are the main focus of comparison.
In any case, good work so far! :-)
-Steve (GSNA R-2306)The Black CabinetA Database of Counterfeit Coins. Attribution Services. NumismetricaCoin Collection Management Software for your Desktop and Mobile Devices.
|
|
Pillar Of The Community
United States
1262 Posts |
Also, now that I remember it, there was an old TPG called "Compugrade" which set off to do the same sort of thing. Unfortunately, the technology didn't exist back then and they soon folded. Who knows? Maybe now's the time? -Steve (GSNA R-2306)The Black CabinetA Database of Counterfeit Coins. Attribution Services. NumismetricaCoin Collection Management Software for your Desktop and Mobile Devices.
|
|
Pillar Of The Community
Australia
6479 Posts |
One of the most powerful ways to detect a fake is to superimpose a picture of the suspected fake over a picture of an equivalent known genuine coin.
If there are differences, they show up very obviously.
A fake detection software of this type could me marketed for the benefit of ALL collectors.
|
|
Valued Member
United States
227 Posts |
Basically SteveCaruso said everything I was going to say (I started writing and didn't see it initially ... lol).
Neural networks are awesome, I took a course in machine vision for my graduate work and loved it... we had some homework to write one for pattern recognition; cool concept. Things get complicated quickly with vision.
Anyway I was also thinking a laser scanner in combination with some type of machine learning algorithm could do it provided you focused on key diagnostics that were really easy to define. The problem is that a lot of coins probably have weakly defined diagnostics even if you think they are "obvious". A lot of things are obvious to us but getting a computer to recognize the difference is like teaching a baby how to do a backflip. So in theory it could work but probably only in certain cases, and it would be difficult even in those cases.
When it comes to replacing humans with machines, you want the machine to be a clearly more effective alternative. In this case the machine would be slower, more expensive, and significantly less reliable than a trained human. And most importantly, at the end of it all you still need the human to grade the coin anyway. So its a really neat idea but not enough reasons to justify it.
|
|
Pillar Of The Community
United States
892 Posts |
Quote:
in combination with some type of machine learning algorithm Sometimes it's helpful to borrow from what we've learned in solving other complex tech problems. The comment made me revisit an algorithm that makes the decision to discard the billions of spam emails every day. There is a remarkable similarity between the method of detecting spam emails and the detection of fake coins and the determination of the coin's specific variety. If a human were filtering the emails it would be a judgment decision based on the content of the email and the person's experience. This is very similar to a human who examines coins to detect fakes. A programming algorithm can't ‘think' so instead it often digitizes the results into a 1 to 100 scale where 100 means ‘absolutely sure' and 0 means ‘ no idea'. Spam software is typically a heuristic algorithm which almost sounds like a contradiction of itself. It makes decisions and then ‘learns' how to make better decisions by repeating itself. Each decision carries a weight depending on how important it is. For example if spam software sees ‘buy Rolex cheap' or ‘offshore pharmacy' in an email it would find that to be adequate proof to immediately assign a 100 rating to the email and delete it. If a fake coin detection program sees a coin of a certain date that should have 13 stars and it only has 12 stars then a score of 100 is immediately assigned because the coin is a fake. This can also apply to determining the variety of a coin. Let's use capped bust half dollars as an example. Several of the years have over 20 different varieties. There are dozens of diagnostics on the obverse like the position of the stars to the milling and on the reverse positions of the ‘A's with respect to the milling. Some diagnostics help us get closer to the answer more quickly than others -- like the alignment of the ‘I' of PLURIBUS to the ‘T' of STATES. This one diagnostic often narrows the number of possible varieties from 20+ down to just a few. This diagnostic would carry a bigger weight factor than some of the minor diagnostics in the algorithm. I often use this web site first before opening my Parsley book. http://www.wix.com/coinzip/bust-hal...tribute/1828OK, back to the problem of worn coins that I raised in my earlier post. Each coin series is different but the major/minor comparison still applies. On capped bust coins the I to T diagnostic is usually available unless the coin is in poor condition. Some of the diagnostics that point to the milling may not be available because of the wear on the edge of the coin. So, a missing diagnostic involving the milling should not trigger a suspicion of a fake coin if other diagnostics can be used to determine the variety. This is a very complex process but -- good news --it's all documented very clearly.
|
|
Topic |
|
Posted 06/12/2012 2:23 pm
