Is AI any good at identifying die varieties?
johnnyb
Posts: 52 ✭✭✭
I keep hearing amazing things about AI (Artificial Intelligence). I think a viable use case for it can be identifying die variety. Imagine what this could do for capped bust halves or other series/date with lots of varieties. Has anyone tried this technology yet for that for that purpose? How is it working?
1
Comments
Let's not confuse "AI" with LLMs like ChatGPT. I'm sure AI tools like those used by NASA or DOD for image analysis could. LLMs are designed around language not images, although they are getting better at images.
All comments reflect the opinion of the author, evn when irrefutably accurate.
Right - the subject of Artificial Intelligence covers a lot of different topics.
Here the relevant one is "Computer Vision", with the subtopic identifying objects in photos (or videos).
https://en.wikipedia.org/wiki/Computer_vision
It would not be possible to do it by "brute force", where you would simply give it thousands of auction photos of bust halves (with no knowledge added from die variety books)
and expect it to identify the set of different varieties and the variety of each photo.
This is because there are not enough auction photos to get enough photos of each different variety to find the identifying die markers.
And the auction photo quality is usually not good enough for the tiny markers.
However, it could be done at some level with a full set of reference photos and known positions of the die markers.
It would be a big effort to do that for bust halves, as it would involve gathering the photos, and adding the locations of the
die markers.
That part alone is similar to writing a die variety book.
Plus there are the basic problems of handling the photo of the coin to be identified.
You need to automatically crop and rotate that photo to align it with the reference photos.
And then have a threshold system to decide if the coin really has a given die marker or not.
And you would still be limited by the photo quality for a given coin that you wanted to identify.
At a less extreme level, one of the members here made a device which can identify wheat cents by date and mint marks to some level of accuracy. It's a similar type of problem, just at a less detailed level.
It is not very fast (15 seconds per coin) but is interesting.
https://forums.collectors.com/discussion/1114794/numi-v3-automated-ai-coin-sorting-prototype
Excellent points. Thank you for such a thoughtful reply.