AI Shows Where It Looks On The Stamp To Identify It. (Franklin)

Say I had a stamp I could not identify. I could scan it and post it here and ask for an opinion. I could ask an AI tool for its opinion. It is the exact same amount of work for me, and I would do the exact same thing in either case - I'd take the opinion and check against references I have access to, and in both cases, I would learn. In neither case would I "trust blindly." We all take for granted the ability to take a high-resolution scan and get highly-accurate opinions from experts across the globe. Does scanning and digital technology take the fun out of collecting? Maybe it does for some. Maybe some collectors prefer to only use a magnifier and see stamps in person to enjoy the hobby.

I think I'm pretty good at identifying tricky U.S. stamps, but I almost never come up the answer by my brilliant self - I have tools and references I know how to use. Usually, my references tell me where to look for secret marks or reliefs or grills or recutting varieties or perforation varieties or color varieties, etc. etc. . .

I bet everyone here would love an AI tool that could differentiate a 594, 596, or 613 from common stamps for the dozens of random people who come here convinced they have a winning lottery ticket. (Actually, that would be trivial to do with no AI and 99.999% accuracy by just having a utility that says "not a rarity" no matter what stamp is inputted.) Anyway - I love this topic. . .and am excited about using AI with my collecting - and I appreciate what imodous has built here!

The heat map might be no more than the AI tool telling us from a mathematical perspective where it,on average, has found the most probable position of the image that identifies the stamp. Maybe, on average, the unique image on a stamp is slightly offset to the right. AI will start looking there, even if you and I find that a dumb algorithm for a stamp with a symmetrical frame and a portrait in the middle.

If you teach the stamp to identify only an equal number of Penny Blacks, Reds, and Blues and Machins, the first range has the portrait in the centre, the second offset to the right. The portrait identifies what range it should look for. Mathematically, the identifying element is offset to the right because of the Machins, but not as far as on the Machins because of the Victorian issues.

Once you know which type of stamp it is, you know where to look for the value.

I think two (or more) things are being conflated here. Being cautious about the capabilities of AI does not make one a technophobe. I love technology. Modern vehicles are a marvel. The amount of computing power in an average car is astounding. I use voice commands constantly in my vehicle and appreciate the smart functions. Scanning images and sharing with the world via the internet is still a marvel.




I am also excited about the possibilities of tech for our hobby.

The problem I have with AI is the way AI is being discussed and marketed. Looking at you MSM. Those throwing it about the most could not really explain how it works. "Now With AI!" is attached to all kinds of consumer goods and services. It seems that the argument is that AI simply appeared and can get rid of those warts and do your homework for you at the same time.

The truth is:

If you cannot see the difference between asking AI and asking people here for opinions, then nothing said here will matter since you think they are identical. They are far from it. Many here know far more than AI is likely to be able to identify.

Strawman much? I did not say AI and human experts are identical. I said the process is the same in that you need to check against references once you get an opinion.

Sometimes. Sometimes not. Depends on the specific stamp and the specific question. Being knowledgeable means that there will be times when the answer is known without looking it up. There is no "one size fits all" answer for every philatelic question.

Two thoughts on the AI...

First, the term AI is currently being bandied about as an advertising/marketing tool much the same way that blockchain was 5-10 years ago. People who didn't really understand how it works and how it can be properly used are throwing it out as a buzzword in an effort to make customers feel as though they are on the "bleeding edge" of technology.

Second, AI can (and will be) extremely useful IF it is properly trained to check the important points for each individual stamp issue. As with any good philatelist, knowing where to look and what to look for are extremely important in discerning which stamp you're looking at. Measuring perforation counts and sizes (original or reperfs), comparing frames of flat or rotary press printings, color variations, etc. all could fall within the scope of what is possible with AI.

I believe I'm in the minority, but as AI matures (and with proper training), it will be as good (or better) at identifying difficult stamps than humans. Just my 2¢.

Phillazilla, firstly, I appreciate you entirely in what you do and know about stamps as is. I agree that the possibilities of adding fun to identifying stamps and a return of "quick id" is no expert guarantee, but those things help the collector.

The LLM is a different model than CNN (Convolutional Neural Network) which is designed for images. Partime, the stamp identifier software is also a CNN, but it is not open source, so to pjr's comment—this is open source software, and I will be making it available on GitHub. NSK and rev collector, while human expertise is invaluable, the model works—I don't know how it works but it does.
Revcollector your trust issues are valid. It could be a tool for experts to find things we didn't know or provide a means to research variations that are difficult to differentiate or specifically for shades of colors if it is specialized to do that but it depends on how the model is trained. I can't speak directly to trying to fool experts but if someone can fool AI then they may be able to fool an expert. In a google search I found that the AI models for stamping were focused on deceit prevention in fraudulent manipulation of stamps and identification of altered stamps. So that may be more of use to an expert.
Rogdcam. It wasn't told where to look. The AI was trained on a large set of images, and it learned how to identify the patterns by itself. It's kind of like how we remember the faces we see. We don't think about how we remember names to faces, we just do.
NSK, about the issues identifying the UK Victorian stamps I suggest creating a model and showing us what it is doing and then reporting it in a topic. Overall as I have started "tuning" my model with more training data I have moreso noticed that the AI is looking at the corners and the perforations but not at the center of the stamp.
I think having these discussions is important to the community and it's great that some of the community is exploring the nuances of how AI is real and how it can be used in stamp collecting.

@imodius




Please, try to read what was written. No issues were mentioned.

I like this discussion. Very thought provoking as evidenced by the various comments.

One thing in my mind is that I would eventually "trust" a tool like this to offer an opinion. However, I would then "verify" the answer using my reference books.

I like this discussion as well and it keeps generating more questions and thoughts.

Given the sheer scope of issued stamps could it ever be economically viable or practical to have a user friendly AI identification tool that works with a very high level of confidence.

Will AI ever be more efficient at stamp tasks than humans are. Is the "setup" required for good output more effort than what an experienced philatelist can accomplish in a set amount of time using simple tools and their cumulative knowledge.

All of this seems to rotate around image quality and parameters. IMO that is the most likely deal breaker. If you cannot make examining a stamp and getting good output from AI a simpler task than manual examination is the juice worth the squeeze.

My opinion is that designing an AI stamp ID tool for every Scott number will never happen because of the money and material access issues. However, I do see a place for AI being used to find alterations and repairs of stamps with values that (possibly) justify the development expense.

An AI tool that could quickly call-out fake coil stamps is an example of the tech being valuable. Again, AI will need very controlled and specific imaging which might include high-res and close-up images. Is this assumption correct?

I would love to know how AI would be able to tell the difference between the natural variations in 19th century paper and actual damage. Especially a professional repair. I have seen rare and high catalog stamps with repairs so well done that even knowing they were there, they were all but invisible. It took a 40x glass to even see them knowing that they were there.

"Sadly, people post the garbage on stamp forums, thus disseminating the garbage and helping ChatGPT believe its own nonsense because it can find it posted on the web."

Garbage in, garbage out. Properly trained by experts, good. Sucking up everything it sees, not so much.

The same way it does with any image classification task—executing the inference stage over a 40X scan of the patient. That part of the answer is easy. The hard part is that the model would have to be trained on a correctly labeled training set of something like 10,000 40X image scans of other such stamps. Sometimes synthetic data can fill out the training set, but the availability of a highly accurate model always depends on the economics of forming or obtaining the training set.

Like all innovations from the invention of the wheel onwards, AI is a mixture of good and bad. In the case of stamp-collecting, it could take some of the fun away, but it could also take some of the drudgery away.

I hope imodius will announce here the release of version 0.1 on GitHub, because I'm looking forward to downloading it, playing with it and perhaps even helping with it.