The model was trained to boost ink and attenuate everything else (on 1r margin). But it is interesting that even the simplest linear regression here (mixing images with certain weights, that's all), produces this good a separation.
I've been looking at 'A' shapes and the first place I looked was the history section on You are not allowed to view links.
Register or
Login to view.
A's by Marci (or his assistant?) appear to be somewhat similar to the leftmost part of the feature on f116v. I have no expertise to compare the writing, but this general shape of 'A' was obviously used some 400 years ago.
[
attachment=9218]
Just for people who can't run the python code themselves, I just want to show that it brings out even a tiny bit more than the image you previously posted. It's really well done.
[
attachment=9219]
Edit: about the A, it is quite similar indeed. If we assume for a moment that this was indeed writing, could it have been A for anno? (And thus perhaps have been followed by numbers).
Edit2: On f17r, the code shows much less of the marginalia than the image Lisa provided though.
If this ' scribble' is compared with the visible scribbles on f86v3 and other places, it appears that these other ones may have been written by a pen, as the lines are thin.
For the case of You are not allowed to view links.
Register or
Login to view. I cannot imagine what kind of writing implement could have created them, apart from perhaps the ones that Koen highlighted in blue. But they are still too thick / heavy.
It would be nice to have MSI scans of these other scribbles, by the way...
(17-09-2024, 08:19 PM)Koen G Wrote: You are not allowed to view links. Register or Login to view.Edit2: On f17r, the code shows much less of the marginalia than the image Lisa provided though.
I noticed that as well. I can think of several possible explanations, without further tests I have no opinion on their relative probability.
- Large blobs of paint on the page affecting the photography via exposure correction/secondary reflection/glow. E.g., there is paint on 17r, while it's absent on 116v and 1r (discounting weirdos). The presence of paint in other areas of the page should not affect the application of the weights, since the weights are applied uniformly. But if at any stage the local spectra were affected by the overall brightness/color of the page, this would change the levels in TIFFs and can boost the background signal. It should be possible to use color charts (or whatever these things lying next to MS are called) to check this, I think, but I have no idea how they work exactly.
- The way vellum was treated or sourced differs, so it has slightly different color overall. If this is the case, training the model on other pages might help. Maybe it's even possible to tell apart different batches of vellum used. Given we have folio 8 in MSIs, maybe we can try comparing it with 1r.
- The composition (or decomposition) of the ink differs.
- Random chance, the background/ink profiles on 1r and 116v just happen to be compatible (this one we can rule out or confirm after processing large enough number of pages).
(17-09-2024, 08:19 PM)Koen G Wrote: You are not allowed to view links. Register or Login to view.Just for people who can't run the python code themselves, I just want to show that it brings out even a tiny bit more than the image you previously posted. It's really well done.
Thank you for posting this! I actually didn't realize until now how much difference there is. This is version 12 of the model, the past images were from version 4. When you just see the small iterative changes with each new version, they don't appear to bring out too much new detail. The main difference between v4 and v12 is the number of extra training points used. Initially I trained models on individual pixels, but then I realized that whatever point I mark as ink/non ink, given the high resolution of the TIFFs, the immediately adjacent pixels in almost all cases should have the same inkness so to say, and they appear to have substantially different values (a lot of variation between adjacent pixels in the images). So I could sample from up to 8 additional pixel values for each manually selected point. In version 12 I settled on using 4 pixel values for each of the manually marked samples - the center pixel and 3 of its immediate neighbors, selected (pseudo)randomly.
(17-09-2024, 01:26 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.Another thing (back to f1r):
Brumbaugh writes in his book (p.115-116):
"After some work with ultraviolet light on folio 1r, my son Robert Conrad Brumbuagh, who was assisting me, proved that indeed someone had recognized the Bacon attribution cipher and had written it here in the margin. It is now badly obliterated and faded. In very small numbers, just above this table, is a date 1 * 3 0, the * illegible. Now, 1630 would fall into the period between the death of Tepenecz (in 1622) and Marchi's inheriting the manuscript (prior to 1644). At first, I thought the table might be earlier, a deliberate invitation to any would-be purchaser to read the ''Bacon'' cipher in the key. But we now favor the 1630 date, and assume that the writer of this table, having found that it read the key text, hoped that it would work for the balance of the cipher manuscript text as well."
Just a thought, maybe the rightmost "a, b, c" in the third column, being offset from the rest of the letters, resembled in UV a separate vertical inscription? 'b' and 'c' could look like 30 together, 'a' would be something that looks not quite like 6, but not totally unlike 6 at the same time, and some random line above them would register as 1?
(18-09-2024, 01:07 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.If this ' scribble' is compared with the visible scribbles on f86v3 and other places, it appears that these other ones may have been written by a pen, as the lines are thin.
For the case of You are not allowed to view links. Register or Login to view. I cannot imagine what kind of writing implement could have created them, apart from perhaps the ones that Koen highlighted in blue. But they are still too thick / heavy.
I think it's possible that this scribble was initially written over a wet spot on the page, which spread around the ink of its bottom part. Maybe after seeing this mess on the page the whole thing was washed off, explaining no visible traces of ink. Then the page was left to dry off and the scribble never repeated (if it was intentional in the first place). Or this scribble was initially thin, but in the process of removing it the bottom part was left moist with water or solvent for too long before rubbing it off, leaving this wide outline. I have no experience in these matters, but intuitively the first possibility seems more plausible, because it explains both why the scribble looks like this and why it was erased, and in the second scenario I would expect the thin lines to persist inside the smudged strokes.
(18-09-2024, 01:07 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.If this ' scribble' is compared with the visible scribbles on f86v3 and other places ...
I am guessing, by the repeated reference to the f86v3 artifact as a "scribble", that perhaps it has not been recognized that the "scribble" seems pretty clearly to be the result of someone writing the same form over and over again. They have written it at least 6 times, and appear to be "practice writing" to perfect the writing of a particular letter, or perhaps it is a pair of letters. (They appear to be perhaps "h" followed by some version of an "f". or "R")
(There is also a large perfectly drawn circle and hand drawn T, identical -- and also the exact same size -- as that in middle of folio f68v1).
BTW, the following is unlikely to work, but worth a try. One can train the same kinds of models on visible light scans by using the ink outlines detected in multispectral images, in an attempt to extract some additional information from the color channels. I don't think linear regression would be able to use this data, but more advanced models could get something, especially if the basic feature is not a pixel, but a group of pixels (ordered or just a "bag of pixels"). I wonder if the TIFFs of the 2014 scans are still available? The link mentioned in You are not allowed to view links.
Register or
Login to view. no longer works for me (404).
Update: I found a link to the stored copies on the Wayback Machine, provided by nablator here <https://www.voynich.ninja/thread-2019-post-44240.html#pid44240>,
but getting the images this way one by one is a bit cumbersome. Turns out there is a bundle with all the images there as well.
Has it been mentioned that the writer corrected (?) part of the right column?
[
attachment=9220]
Additionally, I think it must be noted that the writer omitted "j" from their alphabet. Apparently for English, "j" was added to the alphabet in the early 16th century. I assume that when Marci was writing, his own alphabet likely included the "j". If this is true, then it is clear that he considered the MS older,
and that he was trying to solve it rather than invent his own cipher, because he took into account that it would be without "j".