How Long Before There's A 'face' Recognition Program For Nihonto?

[nagamaki - Franco]

Having recently purchased Markus Sesko's 'Kantei Reference Book, Hamon & Boshi,' it has crossed my thoughts that someone must be at least thinking about this, no? 

 

Thank you.

[Ken-Hawaii]

Isn't that what our faces are supposed to do? Seriously, Franco, I've worked with imaging for most of my half-century career, & I consider a computer-based Nihonto recognition system as highly unlikely.

 

First, just getting any two people to agree on something as basic as the hamon, can be a real chore. What's midare to me, may be notare to the guy sitting next to me. The guys I study blades with are almost never on the same page for jihada, either. If we agreed on sugata, does that hira-zukuri come from Kamakura or shinsakuto? The permutations are as endless as blade condition, & a lens simply doesn't see the same level of detail as the eyeball, anyway.

 

If I had high-resolution images (600 dpi) of every blade that has reached, say, Tokubetsu Hozon, I could probably figure out a way to interpolate (for blades in good condition) at least their jidai & school, but I doubt even that would enable the auto-recognition at the tosho level.

 

It would certainly be an interesting candidate for a deep-learning AI application, but I don't see it coming any time soon.

[Vermithrax16]

I don't get much from oshigata, prefer quality pictures of the actual blades to formulate a memory of different works. I know about zero about AI and computation, but as Ken mentioned, the subtle differences in what is termed what would probably be hard to capture in any algo system based on oshigata.

[Katsujinken]

I actually think the biggest obstacle here is training data: there simply isn’t enough high quality training data (e.g. photos of sufficient quality and detail) to train a model that could approach the accuracy of a trained person – because in truth there aren’t enough swords!

 

So, maybe some day, but as Ken said it’s far, far in the future.

[Surfson]

I think that if Shazam can take in a little music and then identify the artist and song, a good program could at least recognize all swords that have good photos on file.  Now kantei of a blade not in the database seems more unlikely.

[eternal_newbie]

Michael and Ken have the right of it - for this to work as an actual recognition program for providing attributions, you'd not only need a huge amount of data, but have that data be in a standardized format (e.g. all the blades would need to be shot from the same angle, under the same lighting, and in the same condition and polish!). A better usage would be to train it to do things like try to work out the original shape before suriage, or identify flaws that a beginner might miss, such as retempering or reshaped kissaki.

 

You'd probably have more luck with training it to spot gimei (we have AI that scans written signatures for forgeries, so why not chiseled ones) - but I fear this would have a negative impact in the long run since it would give fraudsters the perfect way to hone their art until it passes the machine check, at which point it'd be easy to fool all but the most knowledgeable experts.

[Ken-Hawaii]

Remember that there aren't a lot of blades that have passed two levels of shinsa (I can see a mistake being made at Hozon, but most are likely corrected by TH), compared to the 2 million or so blades that are said to exist. And we'd need that level of identification/authentication to have a base of data that is reliable. Only then would we get to the actual imaging.

 

Rohan, without origami, I think you should assume any signed blade is gimei, & the mei is the very last thing you should look at on a blade.

[IanB]

Years ago, before computers became household items, I tried to compile a book based identification system based on a series of questions such as  'If the blade has XXXX then go to page 45'.  Whist it worked to an extent, the problem the Ken has pointed out meant it was too fallible. You cannot accurately define in words the almost infinite subtle details on which identification is based.

Ian Bottomley 

[CSM101]

Here you can find the next problem:

 

2 Katana, both mumei, both Juyo Token, both with an attribution to Rai Kunitoshi.

Do they look as if they come from the same man?

 

Uwe G.

[vajo]

Uwe, i would say yes. The hada is really similar in my eyes. The first one gets some polishes more than the second. And every good swordmith is able to make every hamon and style. I would say yes.

[Ken-Hawaii]

Excellent illustration of why image recognition > identification is so difficult. The algorithm to identify both blades as Rai Kunitoshi just isn't here yet. Could an AI system be trained to recognize both of them? Sure, but if just one was in the database, the second wouldn't likely be ID'd as the same tosho, jihada notwithstanding.

[nagamaki - Franco]

Hello,

 

Thank you for all of your enjoyable responses. Please continue.

 

Comments in light of things mentioned;

 

One feature that stands out for me when reading Yamanaka's Newsletters revised are those notes which refer to which traditions the smith worked in and the differences seen. While recognizing that there will be significant differences in works by the same smith, as we read Yamanaka, Nihonto Koza, Fujishiro, etc., it becomes apparent that there will be certain identifiable characteristics which make it possible to pinpoint with marked reliability who done it. When writing up kantei discussions for our study group rarely did these references all describe a smith's work in the same exact way. Yet, it seemed at least one reference would nail the description of the sword in question.  And if it can be done in words, it makes me believe it might be possible to be done digitally as well. 

 

Yes, it would take a herculean effort by an organization with access to the best examples available to create a data base with enough depth to cover at least most of the possibilities. Which might be an even greater obstacle in the end to overcome than the tech portion itself. I'm also thinking at this point that this is not to replace the human element, but rather to augment. Computers after all have a knack for picking up and showing small detail that humans might have overlooked or not considered.

 

And, on top of pattern recognition programs where's the current ability to date the age/metal without intrusion at?

 

Thinking about this does make one appreciate even more just what an amazing thing the human mind really is, can be. 

 

Late night ramblings ....

[Ken-Hawaii]

It's fairly simple to assay the metal in a blade, Franco, to several decimal places, & at a reasonable price. But that's not the same as dating it. Radiocarbon dating can be done only on mterial that was once part of a living organism, but since swords are made using coal, charcoal, & other carbon-based fuels, as well as straw to provide the required levels of carbon, dating has been done, but you might want to check out https://journals.uair.arizona.edu/index.php/radiocarbon/article/download/3958/3383to see what the process entails.

[dwmc]

If there were twenty million nihonto collectors/enthusiast willing to spend $99+ usd each for a developed photo nihonto identification program with say a 90 percent probability of accuracy. Maybe it would be worth the cost of development, just maybe. ( Further more, I would try to be one of the first in line.)  Unfortunately....

 

My first PC was a Commodore 64, (1980's)  it came with a couple of floppy disk, one of which was a one line graphic of the planets rotating around the sun, the graphic was equivalent to about what a four year old child could do, possibly not even that good. Look at the advancement in 30 + years..

 

Franco's idea would be great, Ken's thoughts are a bit more of a sobering reality, at least for now.

 

What could possibly be done at this time is a bit like Markus Seskos kantei serious in the form of a desktop application. Also, similar to what Mr. Bottomley had in mind years ago. For instance data input:  Sugata-most like Koto, Shinto, etc,  Nagasa-xxx , Kasane-xxx, Funbari-xxx,  Nakago-Yasurime/patina,  Hamon- most like xxx, basic kantei point questions. Results = X% probability of X.

 

This would obviously need to be undertaken by someone as knowledgeable as Markus Sesko and a company along the line of App. Solutions. Also, what would development cost be vrs potential return?

 

I personally have attempted to kantei several of my Mumei blades by following Markus's kantei series, as well as his e-books which are great. Also, other hard back books as well as online research. Would I have rather had an desktop application to simply input 10 or 12 Kantei points and receive even a 70% probability as to era and 30% probability as to Swordsmith as a place to focus research...Absolutely!

 

Next project Markus???

[Ken-Hawaii]

Sugata = 70% probability on era

Jihada = 60% probability on school

Hamon = 25% probability on tosho

 

And the only application we need is our trained eyes & brains!

[dwmc]

Sugata = 70% probability on era

Jihada = 60% probability on school

Hamon = 25% probability on tosho

 

And the only application we need is our trained eyes & brains!

 

Of course Ken,

 

Only in addition to trained eyes and brains!

 

Dave