Pure Technologies <> Utilizing Computer Vision + Machine Learning for Natural & Artificial Toning

treybenedict

Hey everyone!

Wanted to re-introduce myself, my name is Trey Benedict, I am an Economics & Data Science major at Berkeley who is working alongside engineers from Apple & AWS building Pure Technologies, an S23 Y-Combinator backed startup (Early Investors in Airbnb, Stripe, DoorDash, Instacart, etc).

Our goal is to use technology and software to create transparency within the Numismatics industry. As some of you have seen before, we have already created a pricing engine for all certified coins, but we have also been silently working on color/toning analysis, which I experimented with at the Smithsonian Institution as a visiting researcher earlier this year.

Pure has developed Computer Vision algorithms to generate the following:

1) Analysis & Breakdown of the full-color spectrum on a specific coin using RGB & HSV.
2) Creation of plots of raw color data and eye appeal data.
3) Generative scores of color based on eye appeal (1-70).
4) Detection of artificial color/toning and color progression anomalies.

An example of a naturally toned morgan dollar after being run through the Pure toning analysis:

This received a rating of 70 (one of the higher bounds of color analysis).

An example of two similarly toned morgans, one being AT and one being NT:


Notice the difference in color progression on the bottom plot (HSV). This triggers the algorithm.

A Textile-Toned AT:

Some eye candy...

& a Blast White Morgan for control reference:

The scale for silver/nickel coins goes from:
0-10 Untoned
10-20 Slightly Toned
20-30 Light Toning
30-40 Cresent Rainbows
40-50 Rainbow Toning
50-70 "High-End" Rainbow Toned

Currently, anomalies are detected within color zones of Purple, Blue, and Red, but the algorithm can also detect unnatural patterns since color progression is very similar across silver coins. The model is set to become better with more images over time, however, is statistically significant based on ~200+ tests.

Attractiveness is rated on an eye appeal scale (HSV), which took inspiration from Brandon Kelley's bag-toned Morgan scale. Raw Color is rated upon a scale developed by Doug Kurz (RGB), thanks for putting in the time to research the physics and chemistry behind toned coins!

Also a big thanks to Jacob at Monster Toned Morgans for helping out here as well and hopping on calls and emails to help train the model.

Now I ask you to post a TrueView or high-res of your coin (AT or NT) to help test and train our model as it is adaptive over time with more images! Scores will vary and become more accurate over time as well as the AT/NT training.

Note: We are super busy at the moment but will process requests as they come in over time.

For any inquiries please direct them to treyb@berkeley.edu.

DISCLAIMER THIS IS A BETA TEST, TAKE RESULTS WITH A GRAIN OF SALT

johnny010

Toned coins have a following. I’m sure you could sticker these and create value for the industry.

treybenedict

@johnny010 said:
Toned coins have a following. I’m sure you could sticker these and create value for the industry.

Appreciate the insight! As much as I see viability in charging, I think this technology could be better suited as an educational resource for people to understand more insights about their toned coins.

Have any toned coins? I'll run an analysis for you with a TrueView.

erwindoc

How about some toned Roosevelts?

Project Numismatics

Trey - I read your previous post/announcement with great interest and felt that you were developing a game changer for numismatics. Unfortunately the post disappeared very quickly without a trace. Can you provide an update on your primary effort as well?

TwoKopeiki

Here are a few to test the model. Some of these are not like the others.

johnny010

@treybenedict said:

@johnny010 said:
Toned coins have a following. I’m sure you could sticker these and create value for the industry.

Appreciate the insight! As much as I see viability in charging, I think this technology could be better suited as an educational resource for people to understand more insights about their toned coins.

Have any toned coins? I'll run an analysis for you with a TrueView.

Ok thank you for the offer. I’ll be home this evening and can post one.

treybenedict

@erwindoc said:
How about some toned Roosevelts?

Ran them!





Some pushing boundaries, an updated post coming soon.

treybenedict

@Project Numismatics said:
Trey - I read your previous post/announcement with great interest and felt that you were developing a game changer for numismatics. Unfortunately the post disappeared very quickly without a trace. Can you provide an update on your primary effort as well?

I can provide more info around the time of the Baltimore show (June 8th). Our development has further increased on that side as well.

treybenedict

@TwoKopeiki said:
Here are a few to test the model. Some of these are not like the others.

Ran two of yours before going home from the office.

Lots of Red... Looks like one triggers the algorithm and the other is within natural bounds.
Love these coins!

TwoKopeiki

@treybenedict said:

Ran two of yours before going home from the office.

Lots of Red... Looks like one triggers the algorithm and the other is within natural bounds.
Love these coins!

Thank you. Toned 8 Reales is my passion and I feetl I got a decent grasp on what to look for in terms of color progression as it applies to both fields and legend / devices. I appreciate your analysis and will share backgrounds of these after the results. Coincidentally, the two you tested come from the same collection.

treybenedict

@TwoKopeiki said:

@treybenedict said:

Ran two of yours before going home from the office.

Lots of Red... Looks like one triggers the algorithm and the other is within natural bounds.
Love these coins!

Thank you. Toned 8 Reales is my passion and I feetl I got a decent grasp on what to look for in terms of color progression as it applies to both fields and legend / devices. I appreciate your analysis and will share backgrounds of these after the results. Coincidentally, the two you tested come from the same collection.

To give you a closer insight, the threshold for the bottom coin is 14.5% red for that series according to the images I have pulled of non-toned, toned, and AT. So this is quite literally a coin toss in this case!

Also keep in mind this is a work in progress and only will get better over time!

retirednow

Interesting that u are working off of photos... can you scan directly a slab coin and if so how were the results of using a coin in hand compared to using a photo of the same coin?

logger7

@TwoKopeiki said:

@treybenedict said:

Ran two of yours before going home from the office.

Lots of Red... Looks like one triggers the algorithm and the other is within natural bounds.
Love these coins!

Thank you. Toned 8 Reales is my passion and I feetl I got a decent grasp on what to look for in terms of color progression as it applies to both fields and legend / devices. I appreciate your analysis and will share backgrounds of these after the results. Coincidentally, the two you tested come from the same collection.

Here's an 8 reales, not a lot of blue

treybenedict

@retirednow said:
Interesting that u are working off of photos... can you scan directly a slab coin and if so how were the results of using a coin in hand compared to using a photo of the same coin?

The goal was to be able to scan a coin first and then run a true view analysis for a color score only. I dived a bit deeper one night and came up with an AT algorithm that was on the bounds of statistical significance (<5% error), so I figured I would share.

In the future, we hope to set up a demo portal for non-true view coins to get analyzed as well, but they have to be of a certain quality and lighting.

A quick iPhone image is not high-enough res to be able to do this (unless you have a 14 pro).

logger7

treybenedict

@logger7 said:

Apologies, cannot do copper right now! (working on this but copper/silver/gold tone quite differently)

Although I love this coin's color!

EDIT:

No score set up yet. Here's the color breakdown

david3142

I think this is very cool! I’d be interested to learn more about the algorithm - presumably it isn’t simply looking at percentages but color progressions. From the above seems as though it doesn’t like blue. What does it think of this, which is very much NT:

TwoKopeiki

@treybenedict said:

@TwoKopeiki said:

@treybenedict said:

Ran two of yours before going home from the office.

Lots of Red... Looks like one triggers the algorithm and the other is within natural bounds.
Love these coins!

Thank you. Toned 8 Reales is my passion and I feetl I got a decent grasp on what to look for in terms of color progression as it applies to both fields and legend / devices. I appreciate your analysis and will share backgrounds of these after the results. Coincidentally, the two you tested come from the same collection.

To give you a closer insight, the threshold for the bottom coin is 14.5% red for that series according to the images I have pulled of non-toned, toned, and AT. So this is quite literally a coin toss in this case!

Also keep in mind this is a work in progress and only will get better over time!

That's pretty neat! Would an improperly set white balance / different color temperature affect the accuracy of the analysis? In my experience PCGS photos can often have a slight red shift. Is your training set comprised of PCGS trueviews, or a wide range of images? I would assume there's an easy way to partner with PCGS to get all the images for a series to develop individual color ranges.

Sorry for all the questions - love the concept!

treybenedict

@david3142 said:
I think this is very cool! I’d be interested to learn more about the algorithm - presumably it isn’t simply looking at percentages but color progressions. From the above seems as though it doesn’t like blue. What does it think of this, which is very much NT:

All I can say is WOW nice coin!

It likes these blues
Also, that score is ~10 off from the Moose. One of the higher scores I've seen.

EDIT: I've trained the model on color progression beyond color so it helps to see these great vibrant changes.

treybenedict

@TwoKopeiki said:

@treybenedict said:

@TwoKopeiki said:

@treybenedict said:

Ran two of yours before going home from the office.

Lots of Red... Looks like one triggers the algorithm and the other is within natural bounds.
Love these coins!

Thank you. Toned 8 Reales is my passion and I feetl I got a decent grasp on what to look for in terms of color progression as it applies to both fields and legend / devices. I appreciate your analysis and will share backgrounds of these after the results. Coincidentally, the two you tested come from the same collection.

To give you a closer insight, the threshold for the bottom coin is 14.5% red for that series according to the images I have pulled of non-toned, toned, and AT. So this is quite literally a coin toss in this case!

Also keep in mind this is a work in progress and only will get better over time!

That's pretty neat! Would an improperly set white balance / different color temperature affect the accuracy of the analysis? In my experience PCGS photos can often have a slight red shift. Is your training set comprised of PCGS trueviews, or a wide range of images? I would assume there's an easy way to partner with PCGS to get all the images for a series to develop individual color ranges.

Sorry for all the questions - love the concept!

Yes, thats correct. Color temperature needs to be kept constant for the algorithm to work well. I figured TrueViews are all the same conditions (usually) and would likely be a good place to start. The library I'm using can help suppress in redshift as well as other photography errors which is a great advance of generative AI.

P.S. PCGS you seeing this?

treybenedict

Also if you want your coins to stay anonymous, you can also PM me and I'll generate a report for you.

Mark

@treybenedict

This is fascinating work. Let me give you a few samples. First, maybe an odd ball for you: A heavily toned pattern.

Next a very lightly toned commemorative:

How about a semi-toned commemorative:

Finally a proof dime:

Thanks!!

Davidk7

treybenedict

@Mark said:
@treybenedict

This is fascinating work. Let me give you a few samples. First, maybe an odd ball for you: A heavily toned pattern.

Next a very lightly toned commemorative:

How about a semi-toned commemorative:

Finally a proof dime:

Thanks!!

Sadly do not have a scoring system set up yet for copper, but I can give you a distribution on the color of that coin. Also spectacular!

Will run the rest of yours tomorrow.

Mr_Spud

Cool, how about this one?

And does it work on clad coins? Here’s one I AT’d if you want to try it

Bigtree

Here’s a coin that I’ve received differing opinions on from NGC and PCGS. Would love to know what your algorithm thinks!

TomB

The two Roosevelt dimes you tagged as AT (1947-S and 1962) are NT. I've seen both those coins in-hand and I believe all the photos of the Roosevelts in that post were taken by me using my equipment and processing. The 1947-S has a type of toning and color that is quite characteristic of a particular, accepted, long-term (decades) storage method and I have seen many, many coins with this look.

I'm just sharing this with you to help your algorithm.

treybenedict

Deploying a new version overnight, will be quicker and more simplified results.> @Bigtree said:

Here’s a coin that I’ve received differing opinions on from NGC and PCGS. Would love to know what your algorithm thinks!

I broke down the blue colors for you here. Natural is % of blue that appears on traditional coins that are straight grade and are not AT/QC. Since this one is pushing 7% or so it is on the cusp of statistics which is around 5% the limit making me think its QC.

treybenedict

@TomB said:
The two Roosevelt dimes you tagged as AT (1947-S and 1962) are NT. I've seen both those coins in-hand and I believe all the photos of the Roosevelts in that post were taken by me using my equipment and processing. The 1947-S has a type of toning and color that is quite characteristic of a particular, accepted, long-term (decades) storage method and I have seen many, many coins with this look.

I'm just sharing this with you to help your algorithm.

Thanks! I am going to run a further breakdown on them now. I'm targeting the blue/purple color a bit deeper now and have been getting some good results.

NOTE: I have been basing the algorithmic results based on coins that are straight grade versus QT, so it's not always perfect and could have human-made flaws.

FlyingAl

The idea is great, but I don't know how accurate it could be. Let's try something like this:

ParadimeCoins

Hi Trey,

These are all CAC coins we sold and some with pretty loud colors. Im curious how it will pick them up.

Very interesting program btw, and hope it eventually adds to transparency of NT & AT of toned coins in Numismatics.

Regards,
Shannon

coinsarefun

Very interesting . I have some for you if you would like to run your tests on.

By the way, these are all proofs. So I’m not sure if that interferes with you algorithms?

.

emeraldATV

Does it have to be a True View brand photo for true results ?

ricko

Very interesting. Applying technology to the tarnish progression on coins. Can computer grading be far behind? I think not.... Cheers, RickO

FlyingAl

@coinsarefun

I'll bet that 2011 Silver eagle comes back AT despite the quite attractive natural toning. I think it'll be hard to get the model to accurately show the correct decision on a coin like this, since it's an outlier with large amounts of blue. However, I have no idea how AI learning works.

pursuitofliberty

Have read through this thread several times and am intrigued.

Since your model uses images, what happens with post processing? Most images do not come straight of the lens perfectly, and most top coin photographers do some post processing. Some nefarious people tweak more than others.

I know it was asked, but is your model geared to handle other properly cropped images? Some TV images are not as good as others either, especially in the more recent era it seems.

Here's a few if you want to run them, just because I am curious how it reacts to older not quite MS material

.

.

.

treybenedict

@pursuitofliberty said:
Have read through this thread several times and am intrigued.

Since your model uses images, what happens with post processing? Most images do not come straight of the lens perfectly, and most top coin photographers do some post processing. Some nefarious people tweak more than others.

I know it was asked, but is your model geared to handle other properly cropped images? Some TV images are not as good as others either, especially in the more recent era it seems.

Here's a few if you want to run them, just because I am curious how it reacts to older not quite MS material

.

.

.

Post processing shouldn’t be a problem here. The way I have developed the algorithm runs all images the same way and can adjust for photographers who change their settings. The only difference is that the photos need to be uniform in size and resolution.

I’ll run these for you when I get to the office. On my phone right now.

treybenedict

@ricko said:
Very interesting. Applying technology to the tarnish progression on coins. Can computer grading be far behind? I think not.... Cheers, RickO

Wink, Wink

treybenedict

@emeraldATV said:
Does it have to be a True View brand photo for true results ?

No, I was using TrueViews first because of their uniformity. You can send me a high-res in the thread like others and I’ll generate a report.

treybenedict

@Davidk7 said:

treybenedict

@Mr_Spud said:
Cool, how about this one?

And does it work on clad coins? Here’s one I AT’d if you want to try it

Here you go

gumby1234

Here are some that you can try when you have time.

gumby1234

treybenedict

@FlyingAl said:
The idea is great, but I don't know how accurate it could be. Let's try something like this:

Thanks for the insight! Here you go

treybenedict

@ParadimeCoins said:
Hi Trey,

These are all CAC coins we sold and some with pretty loud colors. Im curious how it will pick them up.

Very interesting program btw, and hope it eventually adds to transparency of NT & AT of toned coins in Numismatics.

Regards,
Shannon

Here you are! What a great group. Can't score the IC but the rest scored very well. Appreciate the post.

treybenedict

@coinsarefun said:
Very interesting . I have some for you if you would like to run your tests on.

By the way, these are all proofs. So I’m not sure if that interferes with you algorithms?

.

Awesome group of coins. All passed according to my tests. High scores in here too!
Note, <5% Statistics level is Acceptable for Blue/Purple Toned coins, so this eagle exhibits the correct blues.

treybenedict

@FlyingAl said:
@coinsarefun

I'll bet that 2011 Silver eagle comes back AT despite the quite attractive natural toning. I think it'll be hard to get the model to accurately show the correct decision on a coin like this, since it's an outlier with large amounts of blue. However, I have no idea how AI learning works.

So for blue-toned coins, I first check to see if there is a large amount of color taking up the coin. Then I apply what they call "masks" in computer vision. Allowing me to segment all the different types of blue (and purple since purple is a hue of blue) that appear, ie light blue, dark blue, navy, etc. I then compare those colors to other "Naturally Toned" blue coins and then give a comparison rating. As you can see above, the breakdown of the blue colors falls into natural and artificial. If the number for artificial is above >5% statistics tell us that the information provided is triggering the algorithm, hence giving my AT/QC basis.

coinsarefun

@treybenedict said:

@coinsarefun said:
Very interesting . I have some for you if you would like to run your tests on.

By the way, these are all proofs. So I’m not sure if that interferes with you algorithms?

>

Awesome group of coins. All passed according to my tests. High scores in here too!
Note, <5% Statistics level is Acceptable for Blue/Purple Toned coins, so this eagle exhibits the correct blues.

.
Awesome and they all are natural. Glad to know it works on proofs. I’ll post some tokens and medals soon.

treybenedict

@coinsarefun said:

@treybenedict said:

@coinsarefun said:
Very interesting . I have some for you if you would like to run your tests on.

By the way, these are all proofs. So I’m not sure if that interferes with you algorithms?

>

Awesome group of coins. All passed according to my tests. High scores in here too!
Note, <5% Statistics level is Acceptable for Blue/Purple Toned coins, so this eagle exhibits the correct blues.

.
Awesome and they all are natural. Glad to know it works on proofs. I’ll post some tokens and medals soon.

Proofs generally give higher scores in my research. I would assume this is because the colors are magnified by the intense die polish, rather than silver luster on mint state coins.

TomB

I'd imagine proof coinage gets higher color grades at least partially because the TrueView images are taken at lighting angles to enhance the color of proof coinage whereas for business strike coins there is likely more of an attempt to extract luster. Therefore, those three-degrees or so of tilt that are required to briefly "flash" explosive color on proof coinage are used to produce more of a glamour shot and, thus, increase apparent color in the images.