Basically, the question is in the title: I've recently purchased a relatively rare coin (only c. 50 known in all ACsearch/Sixbid archives - excluding resales). I've already identified a couple of die-matches (and mine is die-matched also). My methods are fairly simple: I just compare the images visually. It's also doable with a corpus of c. 50 coins. I would like to hear about other methods: how do you check for die-matches?
I adjust the size and superimpose them manually. Like in a thumb cinema . TIF is a master in doing this digitally.
@Julius Germanicus yes, that's what I do as well. However, a few times when I thought "ah, another die match!", I found out that there is a pellet on the one coin that's not on the other, or some other small detail that's different.
I have 2 photographs of different NewStyles that I need to confirm with Thompson if they are new obverses. I am leaving it awhile so that the imprint of the images in my brain have lapsed!
Considering the current sophistication of computer vision, I'm surprised that no one has written a program to do this.
The American Numismatic Society is developing software. Their work is not available to the public. You can watch a long presentation on it at
My son writes computer vision code, so I sent him an email asking him how difficult it would be to develop a program like this. I will report back when I hear from him.
I believe acsearch has an "image search" fonction which work for finding die-matches. I remember there was a discussion here about its effectiveness. https://www.acsearch.info/home.html?nid=17 https://www.cointalk.com/threads/new-acsearch-die-match-previous-sale-function.343249/
@Ed Snible this is very interesting. My PhD is on prediction research, which has some similarities with automatic image recognition. I hope this (ambitious) project will succeed - however, for those that have tried the google image search (i.e. finding similar pictures when feeding google an image) will know that it's tricky. In the meantime, I hope for some less ambitious tricks to perform a die study on a small number of coins.
I've used ACSearch image search quite a bit - it is "hit & miss" - it does sometimes find a die match or two but is very unreliable & to be thorough you still have to trawl through images. Brute force approach: It helps to find a particularly obvious element on your die of interest to scan quickly through large numbers (some quick to see feature or alignment of features). Scanning the images on ACsearch you can zoom the browser view to reduce the need to open up larger images - saving a lot of time.
This thread reminded me that I had planned to write a technical review and walkthrough of the ANS video. I decided I would never finish it so I published my in-progress notes on the tricks and techniques used in the video at http://digitalhn.blogspot.com/2020/08/notes-on-new-developments-in-computer.html .
As a non-technical computer-phobe I wonder how a programme would cope with factors such as die shift, wear, striking differences, Graffiti, countermarks, planchet problems, corrosion et al. You know all the things that can differ individually or together in many combinations between genuine die matches and their appearance today.
How do humans cope with those factors? No one yet knows how well computer programs will deal with all those differences, because the programs have only been tested on a few coins. Yet the early results are impressive. The ANS' CADS program got the correct match in at least one case where the human expert got it wrong. The computer vision approaches merely locate "features", or interesting edges and curves, and count matching features on different specimens. Die shift, graffiti, countermarks, and planchet problems add more "features". If you have 50 coins you can do the die study on your own. If you have 3000 owls the computer can sort them by the number of matches. Suppose the computer says the most similar pair has 80% matching "features". A human can start the die study looking at the best guesses of the computer.
I've tried the acsearch image search function a couple of times, and got nothing. Interestingly, https://ukiyo-e.org/ has an image search tool that allows you to try to find image matches for Japanese woodblock prints (which I used to collect very actively) in their 220,000+-print database. I've used it many times to try to figure out who the artist is, and to identify the print itself, and have been remarkably successful. I suspect the difference is that except for the coloring and possible cropping of the edges, any two copies of a given Japanese print should be identical. Whereas even when there's a double die match, two ancient coins of a given type can look much more different.
Donna, even the same coin is hard to match against itself! With bas relief the light source angle makes a huge difference. The Ex-Numis provenance search only uses the edge of the coin when doing image matching for this reason. Even then, it still doesn't find everything. (Anyone interested in computer aided finding of provenance should watch Dr. Fleuck of Ex-Numis explain the ideas behind his service. ) The Princeton team working on computer die matches constructs a 3D model of the coin. This is discussed in the ANS video ... I have some unpublished tech reports with more data. There approach would be the best if museums and dealers routinely made 3D scans of coins. Yet there are no such images, so Princeton's approach won't work ... yet. Humans can look at a coin and guess where the high points are. No one has done this yet with bas relief coin images. I predict that in the future the data that Princeton has will be valuable for creating a tool guesses a depth map of coins. That depth map will then be used to generate features. For an example of how a computer can guess the depth map of a single image, try https://cvl-demos.cs.nott.ac.uk/vrn/ . If you are feeling ambitious you can take a coin with facing head, crop it to remove the coin border, and submit it to that site. The AI there will make a very freaky 3D of it.
If anyone is interested in developing a program, here's what my son wrote: I don't think it would require any of the new computer vision stuff . . . If you wanted to look into doing it the tool to use would be OpenCV https://opencv.org/
