Automated Card Authentication

[kevinlenane]

So wanted to update the group on my machine vision card authentication project. We've developed some basic algorithms to detect real/fake 1933 Goudey cards based on 7 markers that can detect a real card with extremely high accuracy. As an after-effect it does appear that we can also supply a basic non-biased grade to various components of the card. In short, the problem is very approachable and doable.

The biggest problem I have right now is training data - what I've used (intentionally) thus far have been images taken on my iPhone - since I'd like this to exist in an app. Standard iPhone 10 photos provide plenty of resolution to achieve whats necessary to machine grade - my issue is that I need many images of cards to train the algorithms further and what I've been able to download off the site here is down-compressed significantly and others are in too high resolution to be consistent with the model.

So my question is - how do you think I can best get a high volume and wide variety of 33 Goudey's taken on an iPhone (or even similar quality smartphone). I've considered going into shops and taking pictures of every card I can find but obviously this isnt the most scalable approach. Are there members on here with large raw Goudey collections that have pictures (note that scans aren't necessary but could work depending on the size)? The images are typically 2-3mb in size but I'm just not sure the best way to gather a large volume of unique images. Any advice or suggestions are welcome - no need to post images here since they will likely be compressed in a way that may make them not usable.

Obviously with the card graders shut down the problem is quite timely - so any and all advice is welcome. Thanks in advance.

Kevin Lenane

[kevinlenane]

FWIW - this was the preliminary list of markers we used to start the pipeline - obviously there is more complexity in the actual development process and some additional markers not included around coloring etc. but I thought it might be interesting to see how this particular card set can be evaluated (and also possibly useful as a non-technical document in just evaluating these cards by eye)

https://docs.google.com/document/d/12pFuFxBDYmiBiF9TRwYiH7UDGFxgJPYOO4clLf30Ae4/edit

[LuckyLarry]

I've got a near set of raw 1933 Goudey cards and an iPhone 7. How do I get the pictures to you (text or email)? Do you need pictures of the front and back? Individual pictures of each card or can I line up five or six?

Larry

[BabyRuth]

SGC - Skynet Grading Company

[Rhotchkiss]

VCP has a large database of pictures.

Also, most auction houses should have solid scans. Search under prior auctions.

Then once you get the cards on your screen, screen shot them

[x2drich2000]

Maybe I missed something in one of the other threads, but after reading through the linked document is seems the markers used are unique to the set chosen. Are you going to have to come up with a different set of markers for each set? If so, how are you planning to accomplish this on sets where there is not a huge number of samples to draw from?

[kevinlenane]

It's a good question - my plan here is to start narrow and then expand - the rarer the set the more scant the data for sure. It's something that we'd have to expand set by set starting with the cross section of popularity and value.

I will post a Dropbox link if anyone does want to post picture. I really appreciate all the above advice - I'll try to see if the images on VCP and the existing grading companies are large enough to warrant a web scrape...

[tonyo]

I have high res scans of my former 33 goudey near set on google docs. I can give you access to view and download. But, the scans include multiple cards per scan. Do you need it to be one card per scan/photo ?

[perezfan]

Heritage has extremely high resolution scans as well. They've sold thousands of Goudeys in all grades over the years.

[Case12]

Are you using AI?

[kevinlenane]

Ill take a look at those scans - my email is kevinlenane@gmail.com for Google Docs access. Thank you for the offers - high res scans could work but the best thing is just close up pictures taken on a smart phone and uploaded to the below Google Drive link. There is an example in that google drive folder of the front and back of the cad:

https://drive.google.com/drive/u/0/folders/1Ui7z4ed-K3YJXmv2ciGxCM13GRpy13B_

To explain a bit more on the image size preference - the general idea is to get the software working such that a general consumer can authenticate a card on their own smartphone device. So to create the best machine vision pipeline - the training data should be around that same resolution (which is sufficient for these purposes) Obviously id like to get my hands on some bulk imagery shot on a smart-phone but i will gladly dig through scans that could be similar size and use them as well if thats more what people have here. Alternatively if a site like VCP stores cards at a similar resolution to smart phone pictures the I would run a web scrape and grab a ton of images en masse - so thanks again for all the suggestions.

If someone out there happens to be sitting on a huge pile of raw Goudeys I would be willing to compensate someone for their time taking pictures of them. Obviously im trying to be somewhat economical in the POC stage given the but the real limiter right now is training data. I may need to come up with some creative ways to get this imagery if web scraping doesnt work - maybe compensating photos with actual cards or some other motivational exercise :)

[kevinlenane]

A couple of other answers/clarifications - it does seem like most folks have scans here on Net54 but they seem to be generally of their own collection and are of a huge variety of resolutions. I chose smartphone photos as a base as I would assume they are the most readily available and consistent (in terms of resolution) when it really counts which is during a transaction. This would include a purely web transaction and even more so during any in-person transaction as most would have a smartphone on their person to authenticate live.

On the AI question - right now Im just using machine vision and basic blocking and tackling with no real AI/MI. I can see at some point in the future the system being able to identify other card markers using some form of AI and also potentially learning things from comparison between automated grades and numerical grades that this would give out. It could learn new things and significant relationships between authentication and condition as well. If this ever gets to the investment stage lord knows Ill have to talk about AI a lot :) - but for now its hard enough to just train up the machine vision with screaming kids in the background...

[Jobu]

CardTarget, which is free, also has a lot of scans.

[kevinlenane]

Thanks to all those who replied and reached out - ive managed to collect some imagery but not a ton. Still looking for those collectors with large quantities of Goudey cards/sets that are least in part in raw format. The system will likely work with graded cards as well but the best training data will be of raw images. I will also say that fortunately - for the grading piece - the training data is a bit less crucial and non-set-year specific for the most part.

[JunkyJoe]

It might be challenging, for various reasons, to run the process using digital images (smartphone photos). I think the only fail-proof way to electronically authenticate a card would be to develop a system that uses a 3D scanner to get all sides and edges of the card, with absolutely precise measurements all around.

This would involve the 3D scanning device taking both photos and measurements of the card, and then sending the scanned information to a PC application. The PC application would then process all the measurements and other photo data (common print marks, coloring, fonts, etc) against a database of card attributes for the given set.

Of course, you would need to have the actual card physically available, to be able to place it in the scanner. There are just too many variables involved with using digital images of cards.

[Bigdaddy]

Wouldn't it be cool to have an app on your phone (with a monthly subscription) where you could take a photo or three of a card, send it to a cloud service which would then analyze the card and then send back the type of information that you get from today's submission process?

If phones today can diagnose medical conditions, track your whereabouts in real time and identify people in your photos, then this bridge is not too far. It just needs demand.

[kevinlenane]

Yes this is exactly what I am working on - I'm not going to charge for it though - at least not to start. And the results could be instantaneous - the real challenge is getting the data consistent and accurate enough to be useful

[kevinlenane]

It might be challenging, for various reasons, to run the process using digital images (smartphone photos). I think the only fail-proof way to electronically authenticate a card would be to develop a system that uses a 3D scanner to get all sides and edges of the card, with absolutely precise measurements all around.

This would involve the 3D scanning device taking both photos and measurements of the card, and then sending the scanned information to a PC application. The PC application would then process all the measurements and other photo data (common print marks, coloring, fonts, etc) against a database of card attributes for the given set.

Of course, you would need to have the actual card physically available, to be able to place it in the scanner. There are just too many variables involved with using digital images of cards.
JunkyJoe, you can pick up quite a bit from a basic phone picture including edge chipping etc. It obviously won't be perfect and will give confidence levels but the raw data along w a fingerprint should be pretty useful to some collectors.

[JunkyJoe]

JunkyJoe, you can pick up quite a bit from a basic phone picture including edge chipping etc. It obviously won't be perfect and will give confidence levels but the raw data along w a fingerprint should be pretty useful to some collectors.

I agree there's plenty of good stuff that could be used from a phone picture, especially from a grading perspective, and from an authenticating perspective to a certain degree. For someone wanting a "slab-free" grade, or to verify a card's grade before making a purchase, I think the smartphone app idea would definitely work.

It would also be great for archiving the history of a card as it changes hands through the years (as long as a couple of the sellers subscribe to the app, at least). I think this could be a particularly valuable service of a smartphone-based app. Maybe the app could assign an ID for the card, the first time a card is photographed and graded, and any future transactions/grading/verification could be done using that specific ID for that specific card.

Although, I don't think it could be 100% accurate in terms of authenticating. Especially when dealing with higher-dollar cards, you would still need a human expert (or two, or three) in order to validate the finding of the automated assessment, on various levels. So in the end, if you want a 100% guarantee that a card is authentic, it would need to end up in the hands of grading pros, IMO anyway. Since this would still require shipping the card to a grading company, this is where I think the 3D scanner/desktop computer/database system concept would still be necessary. Imagine the investment $$$ required for a new or existing grading company to integrate this type of a system into their grading process.