Need a collection of photos of Lincoln Cents for a Neural Network

Hello Cointalk!

It has been a while since my last post. I was wondering if anyone had a collection of photos of Lincoln cents in various grades.

About a month ago, I started a project on Reddit entitled CentAutoGrade and the primary focus of the sub is to train a neural network to grade Lincoln cents.

Some grades it is easy to provide data to the NN for such as MS-65 as there are an abundance of photos online. But grades XF and F there are slim pickings (at least in a decent quality).

If someone here already has a collection of cents (wither be graded or ungraded) any and all data would be greatly appreciated!

You can PM me here and I will send you my email address, please only do so if you have the time and are willing to share. This project is intended to be completely open source and is being developed on volunteer time. If anyone here is willing to contribute to the project in any manner, feel free to contact me as well.

And above all else, happy turkey day!

 

Log in or Sign up to hide this ad.

Don't you try to fool me! Your gonna sell your fancy neuron network to PCGS and NGC and then a computer will grade my coins. Conspiracy theory proven!

 

I don’t know if you’ve already investigated it, but PCGS’s CoinFacts does have some of their TrueViews uploaded for all of the dates and mintmarks. Scarcer dates like 1909-S VBD and 1914-D have a pretty good spread of grades (even lower grades). All of the images are of good quality, so they should be good training images for a neural network.

CoinFacts used to have all TrueViews uploaded, but unfortunately unscrupulous crack-out artist dealers (i.e.- dealers who crack out and resubmit coins for higher grades) were offended that CoinFacts listed all TrueViews for a coin (including TrueViews that were taken when the coin was graded lower). Having lower graded TrueViews available for viewing on CoinFacts interfered with their business practices, so they complained to PCGS and lobbied them to removed ALL TrueViews from CoinFacts. Fortunately PCGS eventually came up with a compromise where they uploaded TrueViews for coins that were listed in registry sets. While not a perfect solution, this nerfed version of CoinFacts is all the better it’s going to get.

 

Hahaha where your conspiracy has fault is that I'm actually the owner of NGC /s

To NSP, I have already stripped as many photos as I could from PCGS and I got around 20 of each grade. But I believe the model will benefit significantly from more data. While I am no computer science major (not am I a major at all) I believe the more data I train the model on the more successful it will be in the long run.

 

Your the owner of NGC? huh?

 

@BadThad

 

I honestly wish dear stranger. Sadly I am just a broke nuismatic.

 

I would contact Heritage and Great Collections and ask permission to use their auction archives. They've sold thousands of cents. You won't find a better (or more consistent) resource.

 

Not a bad idea, I can be a little charismatic. But I'm wondering if they would do such a thing! Worth a shot in the dark I suppose (Cited from: https://www.cointalk.com/threads/ne...of-lincoln-cents-for-a-neural-network.350818/)

 

They are usually pretty generous. I've worked with Heritage before to use pictures in an article. They just wanted me to cite the photos as being theirs (a perfectly reasonable request).

 

And you want us to do your research........crack the books my friend.

 

Most people would define it as using your assets to their highest potential.

 

Heritage, Stacks/Bowers, Great Collections. You should be able to find lots of 09-S VDB And 14-D in lower grades. Stay away from 22, because those will cause problems with your training.

You will need to use photos from more than one source or you will have ground truth data that is corrupted by how photos from a single source tend to turn out. Another thing you should consider is using a high-pass filter on the images to remove or reduce the luster or sheen in the photos of the coins, as these are prominent features in photos that you don't want influencing the training. At the time of inferencing, you would apply the same filter to your test data.

The problem is that you have to select enough photos of each grade that the network can differentiate the grades. Luster and very small features can make a big difference in these grades. The source of your test data will be even more important here, since variations in photo technique can have a large impact on the images. I think you'll have better luck with circulated grades.

Good luck, and let us know how this works out.

 

Last I heard , PCGS, about 10 + years ago acquired most of the algorithms and "facial " recognition pattern analysis related to 'objects' . Those used to be on the internet patent searches, but you have probably already found them as at first most collectors thought they were going to begin to use them for grading, but I think they only used them for specific data points for identification points to trace high end coins if lost or stolen and similar one came on the market ( Although I have nothing to prove that but hearsay). However, in the market the parent company doesn't mind spending a lot of money to protect what they feel is possibly of good value to themselves or might compete with their products. IMO Jim

 

Doug, the entire point of machine learning and training a neural network, as the OP is trying to do, is that it needs multiple inputs to learn. It *needs* a variety of data, so that it can be reliable across a broad spectrum of coins. If you only trained it on Heritage pictures, it would not be able to grade anything else.

Your other quibbles are red herrings. Computer grading 20 or 30 years ago when they first tried it wasn't accurate because the technology just wasn't there. But, I'm fairly certain that it would be achievable now.

 

The neural nets look for correlations. They really aren’t very bright. But the more data they have the deeper the correlations that can be found.

Doug actually mentioned several.

• All Heritage photos are similar.
• Older holders may have a lower grade bias.
• NGC and PCGS and ANACS may have biases.
• TrueViews are glamour shots.

The software may not know any of this they way we do, but if this metadata (TPG, grade, photographer, year graded, coin denomination, year, mint mark) is input with each photo, the correlation will be mathematically derived and weighted.

While the experiment may fail, it might just quantity what experience already hints at. If we adjust for the discovered biases we may come out the other side with a repeatable model that has predictive power. Perhaps it won’t give a universal definitive grade, but it might give a prediction of what each TPG would grade it.

 

I'm responding to two entirely different parts of the OP.

OP starts off asking for circulated coin photos. The first comment is referring to gloss or sheen seen on circulated coins where luster doesn't come into play for the grades. I provide a solution for this. A neural network may learn that this is part of the solution, but providing filtered data will allow it to "concentrate" on more important features.

Later, the OP talks about plentiful pictures in high grade, about which I caution that the luster on the uncirculated coins can be a major player in the grade, and that differences in photo techniques can be seen as more prominent features from those that are relevant for grading.

This is the subject of a debate I'd like to take part in at a major coin show -- "The Future of AI in Coin Grading." I feel qualified to take both sides here -- how it can't work and how it can work.