(23-09-2024, 05:48 AM)oshfdk Wrote: You are not allowed to view links. Register or Login to view.I can try training a model on the shape you give for the top right corner of 1r, on MSI TIFFS there is also something that looks like a small inscription a bit to the left of it and something that looks like another small inscription at the top of the page, and another shape that looks like a small letter. I'll create a separate thread for this if you don't mind.
Thank you, oshfdk! It would be great!
Oshfdk, would it be possible to bring out the pen lines on a regular page in a similar way? It would be extremely handy to run a python code and get black and white outlines of whatever pen work is detected. Also for the outlines of the images etc.
(24-09-2024, 12:31 AM)Koen G Wrote: You are not allowed to view links. Register or Login to view.Oshfdk, would it be possible to bring out the pen lines on a regular page in a similar way? It would be extremely handy to run a python code and get black and white outlines of whatever pen work is detected. Also for the outlines of the images etc.
I have experimented a bit with training a model on visible light scans using point sets taken from MSIs (trying to restore from visible scans as much information available on MSIs as possible). Tentatively, this produces a model that is not bad at separating different layers on the existing high resolution scans. There are two caveats though: resolution will suffer somewhat, since it takes a group of pixels (2x2 or 3x3) to have enough statistical data for more or less satisfactory separation and the separation is not linear, meaning the process will be much more computationally demanding, it's no longer just about summing a bunch of layers with weights. This probably won't work on JPEGs, only on the original TIFFS, because the compression of TIFFs into JPEGs works by removing exactly the extra information the model uses. The algorithm will still treat each small area of 2x2 or 3x3 pixels separately from the rest of the page, so there is some confidence that it won't hallucinate things that are not there.
Maybe if I stumble upon a model that works reasonably well, I will just generate separate sets of layer images for each page and share them.
On a different but related topic, I have been experimenting with a method that removes the paint from images, leaving only the ink used for writing and line drawing.
This seems straightforward, but that turned out to be a misleading first impression,
In any case, it is working reasonably well.
Do you think that your algorithm could separate the two (writing vs. drawing)?
(24-09-2024, 07:50 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.Do you think that your algorithm could separate the two (writing vs. drawing)?
Just to check that I understand it correctly, we are talking about separating the writing ink and paint, not the text ink and ink outlines of drawing (which most likely are physically indistinguishable), is this right? And we are talking about visible light scans, not MSIs?
It's hard to tell without actually trying this out. This looks possible. It should help that pages included in MSI imaging had various kinds of paint, so can be used to automatically mark training areas.
(24-09-2024, 07:00 AM)oshfdk Wrote: You are not allowed to view links. Register or Login to view.Maybe if I stumble upon a model that works reasonably well, I will just generate separate sets of layer images for each page and share them.
That would be cool. Not even in the first place to glance new insights (those will be limited from these scans) but just to have a relatively clean version of the line drawings. And a higher contrast version of the text.
(24-09-2024, 09:15 AM)oshfdk Wrote: You are not allowed to view links. Register or Login to view.Just to check that I understand it correctly, we are talking about separating the writing ink and paint, not the text ink and ink outlines of drawing (which most likely are physically indistinguishable),
Indeed, I meant separating the writing in ink from the outline drawings in ink.
You may assume that the paint has already been removed.
Physically, they appear to be the same, but of course the shapes are different. I was just wondering if your algorithm can take the shapes into account....
I can do it to some extent, as I have a 'box' definition (coordinates) for each word in the RF transliteration. These boxes are slanted along the direction of writing. However, this is not perfect.
See You are not allowed to view links.
Register or
Login to view..
(24-09-2024, 02:00 PM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.Indeed, I meant separating the writing in ink from the outline drawings in ink.
You may assume that the paint has already been removed.
Physically, they appear to be the same, but of course the shapes are different. I was just wondering if your algorithm can take the shapes into account....
I can do it to some extent, as I have a 'box' definition (coordinates) for each word in the RF transliteration. These boxes are slanted along the direction of writing. However, this is not perfect.
See You are not allowed to view links. Register or Login to view..
This looks cool. I think, to automate this without relying on external coordinates would require a different kind of algorithm entirely. Basically, this would be almost the same as creating Voynichese OCR, because once an algorithm can reliably tell apart Voynichese from everything else it's just a few steps from automatically parsing Voynichese. I'm not sure I have this much free time on my hands right now
I've tried to refine the shape of the squiggle by training a separate model on the dark parts of the squiggle itself.
[
attachment=9287]
It's still very blurry and the shape very uncertain, but there are a few possible new interpretations, if we consider this as a signature. First of all, after reviewing a lot of European signatures of the XV-XVII centuries and looking at this new image of this squiggle, I'm no longer sure the first letter is 'A'. I haven't seen a single case of this kind of tall and extremely thin 'A' with no visible horizontal bar in signatures, while I've seen the same kind of shape used a lot in name initial Ls and some other letters, where the leftmost stroke is just a decorative flourish. It's also possible that this leftmost stroke actually goes farther to the left and curls up, the dark spot to the left consistently appears together with the rest of the squiggle when using most models.
There is also something that looks like a loop over the right part of the squiggle, it's hard to tell whether (a bit more likely to me) the loop starts from the right bottom of the squiggle and continues up to the top of 'A-or-not-A' shape (and maybe beyond), or vice versa, starts at the top of 'A-not-A' and goes to the right bottom. What's interesting, this loop goes over one of the big wormholes in the vellum, which could indicate that it's older than the wormholes, and also it looks like the loop was made when the vellum was folded a bit along the bottom right fold, that this loop seems to cross. At least the part of the loop to the right of the fold is a bit offset vertically from the rest of the loop.
Overall even if this is an actual signature, there are many possible ways to interpret it based on the preserved parts. E.g.,
[
attachment=9291]
So, without some sample signature of a potential past owner (or author/scribe) of the manuscript I think it's nearly impossible to get any information from this.
I've also tried to apply the same model to other pages, to see which elements it will highlight, looks like it's most sensitive to text outlines (maybe this means that it's sensitive to locations where the ink is only present in tiny amounts, I'm not sure).
[
attachment=9288]
It's weird how it gives the nymphs a frown
