AI grading 0

[Bignubnumismatics]

 

A friend has been working on a website and an AI grading tool for quite some time now. It has come a long way from where it started, but it still needs revisions and help from the public. If you get a chance, please view the AI grading tool and provide some feedback. We're hoping to make this tool, and the website, accurate and mainstream enough to help beginning and advanced collectors. Any help/feedback is welcome and appreciated. https://coinshow.us/ai-grader

(I would recommend opening it on a PC, he hasn’t prepared it for mobile use yet. Also, it’ll make you create an account so that you can upload images)

[bsshog40]

Might be a tad hard to get someone to open a link on here from a person with 1 post! 

[gmarguli]

And all these years I've been foolishly rotating coins in the light when I grade them. All I needed was a flat picture to grade them. 

[RWB]

A few visual samples and related conversation might allay concerns about unknown links. Tell us something about the project; what works what does not work; how AI is better than "Dumb-n-Dumber;" etc.

Edited November 27, 2020 by RWB

[Hoghead515]

3 hours ago, bsshog40 said:

Might be a tad hard to get someone to open a link on here from a person with 1 post! 

Something smells fishy. I seen another post kinda like that earlier on another topic. 

[Bignubnumismatics]

I've posted on here before with Barber dime varieties, I have the same username just opened a new account with a new email.

 

[Bignubnumismatics]

 

I just had spaces in my last username.

[Jonas Denenberg]

2 hours ago, RWB said:

A few visual samples and related conversation might allay concerns about unknown links. Tell us something about the project; what works what does not work; how AI is better than "Dumb-n-Dumber;" etc.

Hi, I've never used the NGC discussion board before (until now, to reply to you) but I made AI grader. Regarding your first point, I realize that it probably a *bit* sketchy that someone like @Bignubnumismatics here with a brand new account is posting a link on the NGC chatboard but I can assure you that our site is 100% secure. The only info we ask for is your email, so that you can create an account and upload images. When it comes to passwords, we encrypt everything using the bcrypt (https://en.wikipedia.org/wiki/Bcrypt) algorithm, which means that everything is so secure that even I wouldn't be able to figure out your password, even if I wanted to. Your password is stored on our servers as a cryptic string of 70 letters that would take billions of years to decode. The reason you have to create an account is so that we can process your images, and so that we can save your grading data so that you can look at it later (anything can be deleted with the click of a button if you don't want us to store your data). Second, about the project, here is what I have going so far:

- Deep Cameo Proof/Cameo Proof Detection

- Strong/weak strike detection (Not very accurate as of now)

- Contact Mark detection

- Toning detection for silver coins (Using the Sunnywood https://coinshow.us/sunnywood System)

- Color grading for copper coins

- Corrosion detection for copper coins

- Patina/Luster detection

It also, obviously, determines the grade and eye appeal of your coin. Right now, the grader typically gets within 10-15% of the coin's actual grade, so it's more of an experiment as of now. You can provide feedback each time you grade a coin, which is used to improve the accuracy of coins that you grade in the future. I wish this wasn't necessary, but to support images of any quality, it is necessary. My goal is to get it down to 5% in the future.

I've attached an image created by the contact mark detection algorithm, which is something I just added to the site yesterday. (I make updates frequently 😁 ). I also attached an image created by the toning detection algorithm. (If it looks confusing, it's using the sunnywood system https://coinshow.us/sunnywood)

[Jonas Denenberg]

1 hour ago, Hoghead515 said:

Something smells fishy. I seen another post kinda like that earlier on another topic. 

It might have been referring to another site that's also making an AI grader. Their grader hasn't been released yet though. I'm the one who made the AI grader, by the way. I just made an account to respond to people on this thread :)

[bsshog40]

Yea I found a couple others that posted ATS. Link is ok!

[RWB]

A couple of casual comments with regard to criteria.

1. - Deep Cameo Proof/Cameo Proof Detection
2. - Strong/weak strike detection (Not very accurate as of now)
3. - Contact Mark detection
4. - Toning detection for silver coins (Using the Sunnywood https://coinshow.us/sunnywood System)
5. - Color grading for copper coins
6. - Corrosion detection for copper coins
7. - Patina/Luster detection

#1 and #3 have been discussed here and elsewhere many times. Commercial SW can do these extremely well. For #3, "too well" might be the better description; good results depend on binning algorithms selection/designation. #2 is tough because there are few really uniform prototypes, coins are not produced with precision of detail as a goal - unlike industrial and IT components. Even the best detailed coins do not match sculpted designs, and modern digital technology almost breaks as much as it mends.

#4, #5, and #6 are subjective or depend on standards not used or understood by coin collectors. "These are centered on ASTM E308: Practice for computing the color of objects using the CIE System, provides a full description of calculation of CIE tristimulus color scales along with definition of CIE illuminant and observer tables." They are possible, but will take a lot of simplification for coin collector use.

#7 "Patina and Luster"are different things with entirely different causes, so I don't understand that one.

Edited November 28, 2020 by RWB

[Insider]

I predicted something more useful along the same lines as your AI Grader.  

IMO, something much more useful that you could charge a membership fee for would be AI (Foreign/ancient coin) IDENTIFIER!

Download an image and find out what the coin/token is.  I predict a TPGS may do this eventually now that they are imaging stuff.

 

HOW ABOUT IT NGC?

Edited November 28, 2020 by Insider

[gmarguli]

11 minutes ago, Insider said:

IMO, something much more useful that you could charge a membership fee for would be AI (Foreign/ancient coin) IDENTIFIER!

Download an image and find out what the coin/token is.  I predict a TPGS may do this eventually now that they are imaging stuff.

There is Coinoscope for your phone which will do something like this. I've tried it a few times and it is hit/miss. It nailed a hard one, but missed easy ones. 

You can also search Bing/Google using an image and it will find the coin for you. 

[Bignubnumismatics]

Large cent varieties are next on the list , but that will probably take quite a while.

Edited November 28, 2020 by Bignubnumismatics

[physics-fan3.14]

I recently watched a video about this. Not sure if this is the same website as the OP in this thread is referring to. It is quite basic so far, but shows some interesting ideas. 

 

[RWB]

The video posted above is interesting, but seems to be trying to do too much  and require a lot of user intervention. It's also way behind commercial recognition systems such as those used by TI and other technology companies. Suggestion: Start simple with fixed parameters of coins for basic geometry, then concentrate on differences not similarities. Use only unslabbed coins to develop the system parameters.

In any event, I encourage the developers to keep trying.

[messydesk]

AI grading is an interesting exercise, because it shows just how much you need to understand something to build an AI system that can do it well.  This is a good example of a classifier problem, where the classifications are the grades of a coin.  A standard approach to this is to throw a lot of image data together with ground truth grades (I know what you're thinking) and let the system learn how to classify unknown images.  This will work pretty well with wholesome, circulated coins with flat appearances that can be graded pretty much as line drawings and imaged on a flatbed scanner with repeatable results.

Uncirculated coins pose a different challenge.  If you look at how people grade them, there is a lot of tipping and twirling of a coin in the light to get a full picture of what it looks like.  If you translate that process to a computer, that is an incredible amount of data that needs to be acquired, and acquired without loss due to an insufficient imaging setup.  We've all seen countless GTG threads everywhere that have unexpected grades associated with a coin.  We are grading these with one image.  Sometimes they're good images, sometimes not, but it's always that case that a single image isn't capturing everything about a coin. More nefariously, it's sometimes the case that a single image is hiding specific things about a coin, such as a patch of hairlines, a bad hit, or a scratch.  If I'm trying to rip of eBay buyers and I normally use bad (for example, overexposed) pictures of problem coins to do this, knowing there's an AI grader in play, I might decide to try and game that with my photos.

To me, what would be the most impressive is a data acquisition system that could acquire the data necessary to train such a system well and then to actually perform the grading on test coins.  From there, the next step would be to see how to optimize the whole process.

Simpler AI tasks related to coins include identification, coarse grained attribution (think Overton varieties, not VAMs), and maybe even AT/NT assessment.

 

[RWB]

A primary impediment is mixing data measurement with subjective opinion. An AI system has no "opinion"  - only an aggregate that imitates opinion to some degree of human detail. An efficient AI system will be compact; it "asks" the right questions....Few do that. It is too "easy" to scale up clutter from Terascale to Petascale flops, then issue a Press Release claiming some sort of AI "breakthrough."

In system development data collection and opinion imposition are in opposition, so must be separated and analyzed individually. This follows my initial comment about patina and luster, etc., above.