19-10-2025, 04:54 PM # USED (This post was last modified: 19-10-2025, 05:01 PM by # USED Jorge_Stolfi. Edit Reason: Fighting the editor about blank lines # USED ) Like a responsible parent, I have been trying to pass on to my # USED computer my Superior Pareidolia skills. Specifically, the ability # USED to see inked details that were painted over. # USED # USED The Painter who applied the semi-opaque tempera colors often # USED painted over inked outlines. Examples are easily seen where these # USED inked strokes were still dark and clear, like (A,B) below. # USED # USED Besides obscuring those strokes, it seems that the painting also # USED washed away some of the ink, and sometimes deposited it a short # USED distance away, as in (D). # USED # USED Thus any ink strokes that were already quite faint and faded, # USED like (E), must have become invisible to the naked after being # USED painted over. And that is why we need Artificial Superior # USED Pareidolia. # USED The idea is as follows. # USED Take an image of an area which is suspected of having "invisible" # USED drawings or text under some semi-opaque paint. Select a set of # USED pixels A representative of what one wants to detect, like places # USED where there is definitely ink covered by green paint. Select one # USED or more additional sets B, C, ... that are to be distinguished # USED from A -- like places where there is green paint with but almost # USED surely without ink underneath. Look at the colors of those pixels # USED as points of three-dimensional space, within the unit cube where # USED (0,0,0) is black, (1,1,1) is white, (1,0,0) is red, etc. Here is # USED an example with three subsets of a page, representative of blank # USED vellum (red), dark text ink (green), and green paint over blank # USED vellum (blue): # USED # USED Approximate each cloud ou points A, B, C, ... by a trivariate # USED Gaussian probability density function (PDF). This can be # USED visualized as a fuzzy ellipsoid with varied dimensions along # USED three axes, with some generic orientation in space. Take each # USED pixel of the image and use Bayes's formula to estimate the # USED probability that the pixel belongs to each distribution A, B, C, # USED ... or is an "outlier" that probably does not belong to any of # USED them. Write one grayscale image for each set, showing the # USED probability of each pixel belonging to that set. # USED # USED # USED Ideally we should do this with high-resolution uncompressed # USED multispectral images with frontal illumination and linear encoding. But # USED we don't have multispectral scans for any of the pages that may have # USED significant details hidden under the paint. (The herbal pages have green # USED paint, but the ink that can be seen under it is just boring nervures or # USED leaf outlines. At best, those images could be useful to validate this # USED approach.) And even those that we do have are taken with oblique # USED lighting that creates light and dark spots at every tiny bump on the # USED vellum surface. # USED # USED So we must do with the Beinecke 2014 scans, which have frustratingly low # USED resolution (some ink traces being only a couple of pixels across), only # USED the three RGB color coordinates, oblique illumination, non-linear # USED "gamma" encoding, and complex JPEG compression artifacts. But, sigh, # USED that is life... # USED # USED [To be continued] # USED # USED All the best, --stolfi # USED # USED [Sorry for the big images, but I couldn't figure out how to # USED insert only a thumbnail of the attachment, with the full version # USED opening on a click. Is that possible?] # USED ###################################################################### # USED 20-10-2025, 03:40 AM (19-10-2025, 06:16 PM)oshfdk Wrote: This is # USED very similar to what I tried last year, [...] The full gallery: # USED https://oshfdkbw.pages.dev # USED # USED Thanks! I think I am getting slightly better results for "invisible" ink. # USED For instance I get a stronger signal for something between the feet of that nymph, and maybe for something under the east end of the "backrest" she is leaning onto. We will see... # USED ###################################################################### # USED 20-10-2025, 10:44 AM (This post was last modified: 20-10-2025, 10:47 AM by Jorge_Stolfi.) # USED (20-10-2025, 07:24 AM)oshfdk Wrote: # USED g of the parameters of models one can see/reveal almost anything. # USED # USED Indeed. But this approach is at honest at least in the sense that the # USED final classification is made independently for each pixel, based only on # USED its color; without trying to look for multi-pixel patterns like lines or # USED characters. Which is where actual pareidolia comes in. It will be left # USED to the human user to "see" such patterns on the computed probability # USED maps. The user's choice of sample pixels will influence the # USED classification, but only through their colors, not through their # USED positions or adjacency relations. # USED # USED Quote: I suppose if a model is trained on this folio it's possible to # USED get much more curios results, but I was actually aiming for maximum # USED certainty, so I only trained on samples from 6-7 folios and then applied # USED the result to the whole MS. # USED # USED I think that training separately for each page (or even for each clip # USED from each page) is justified because there seems to be some overall # USED variation from folio to folio in the colors of parchment, ink, and # USED paint. For instance, the green paint of f8r seems to be more bluish than # USED that used in the Bio section. # USED # USED And there is also a much larger variety of paints and stains over the # USED whole MS. On this particular image I don't have to include the red, # USED blue, yellow, and rusty paints as separate provinces. They will be # USED classified as "OTHER" without significantly affecting the classification # USED of the other provinces of interest. And I don't have to worry about # USED ketchup stains, or the gray offset from blue flowers, that are important # USED "noise" features on some other pages. # USED # USED All the best, --stolfi # USED ###################################################################### # USED 20-10-2025, 12:16 PM # USED (20-10-2025, 10:57 AM)oshfdk Wrote: # USED # USED it helps if the result can be independently reproduced via a simple # USED (ideally linear, or maybe polynomial) combination of channels. # USED # USED That will not be the case, because Bayesian classification with Gaussian # USED distributions is inherently non-linear, and usually extremely so. # USED # USED For instance, imagine that you have only two Gaussian classes (plus # USED "OTHER"), where class A has a very broad spherical distribution centered # USED at middle gray (0.5,0.5,0.5) and class B has a much narrower one # USED centered at slightly darker gray (0.4,0.4,0.4). Bayesian classification # USED will assign class A to colors inside the A sphere, except within a small # USED region around the darker gray, where it will say B. A linear classifier # USED will be unable to delimit even the A sphere, much less the B hole inside # USED it. # USED # USED That is why linear vector classifiers are usually applied to non-linear # USED functions of the inputs, the (improperly) so called "kernels". Which # USED requires the user to come up with suitable kernels. If one tries to use # USED as kernels all polynomials on the input coordinates up to degree (say) # USED 4, one gets so many kernels that the classification will probably be # USED garbage. # USED # USED That is also a danger if one uses "magical" non-linear classifiers with # USED zillions of internal parameters, like neural networks... # USED # USED All the best, --stolfi # USED ###################################################################### # USED 20-10-2025, 02:18 PM (20-10-2025, 12:31 PM)oshfdk Wrote: So, if there is # USED (R,G,B) -> (Ink, Paint, Vellum) (or whatever it is detecting), which # USED always produces the same result for the same R, G, B and the result # USED doesn't change a lot visually if I replace all original RGBs with (R + # USED r1, G + r2, B + r3) were r1, r2, r3 are reasonably small random # USED integers, say [-2, 2], then I'd say the whole pipeline looks reasonable. # USED # USED The report I just posted shows the results of that basic analysis. What # USED do you think of them? # USED # USED That test still does not try to detect the ink-under-paint "province", # USED which is my ultimate goal. I have tried the latter already (with seven # USED distinct color classes plus "OTHER") and the results are encouraging, # USED although I have still not found the best subsets and the best way to # USED combine the results (like ORing the "dark ink under green paint" map # USED with the "faint ink under green paint" one, and things like that). # USED Please stay tuned... # USED # USED All the best, --stolfi # USED ###################################################################### # USED 20-10-2025, 06:04 PM (20-10-2025, 05:32 PM)oshfdk Wrote: Where do we get # USED this prior probability, is it just by estimating the proportion of the # USED pixels representing paint, ink or vellum? # USED # USED That is the fundamental limitation of Bayesian inference: it does not # USED give the probabilities of each possible cause, it only tell us how to # USED change out a priori probabilities. So you must specify those priors in # USED order to use the formula. # USED # USED When the evidence is strong enough, the results do not depend much on # USED the priors. These matter only when the evidence is ambiguous. In the # USED worst case, when the evidence carries no information about the cause, # USED the results will be the same as the priors. # USED # USED When it rains at night, streets are usually wet by the next morning. If # USED you have no special information about the night's weather, and you see # USED the street wet in the morning, it is reasonable to assume that it # USED rained. Bayes's formula will say so, whether your prior for "it rained" # USED is 1% or 99%. # USED # USED That's because P(Wet|NoRain) is very small; say, 0.0001 (but not zero, # USED because there may have been a flood or a street washing truck mat have # USED showed up. While P(Wet|Rain) is basically 1. Therefore # USED P(Rain)P(Wet|Rain) is still much bigger than P(NoRain)P(Wet|NoRain) in # USED any case, and these numbers will become ~1 and ~0 after normalization. # USED # USED But if you are sure that it did not rain, because you have been out # USED stargazing in the garden all night, the wet street should not make you # USED change your belief. That's because your P(Rain) will be essentially # USED zero, and then P(Rain)P(Wet|Rain) will become much smaller than # USED P(NoRain)P(Wet|NoRain), even if P(Wet|NoRain) is only 0.00001. # USED # USED In the analysis reported in that document, the prior probability of # USED "OTHER" was set arbitrarily at 0.05, and that of each of the three other # USED classes was set to 0.95/3 = ~0.32. # USED # USED All the best, --stolfi # USED ###################################################################### # USED 20-10-2025, 07:18 PM # USED # USED This image (a clip of f79r magnified 400%) shows # USED some of the challenges on the way to uncovering drawings that have been # USED painted over. # USED # USED There are some ink traces, like (A), that were quite dark to begin with # USED and suffered little from being painted over, apart from their color # USED getting mixed with the paint color. # USED # USED But there are some ink traces, like (B), that are very faint, almost # USED invisible. Would we be able to recover them if they were painted over? I # USED think that smudges at ( C) look like they were the original outline of # USED the nymph's left foot which was incorrectly retraced as the thick dark # USED stroke a bit further to the NW. Those smudges are only a little fainter # USED than the toes of that same foot, which almost certainly were there. # USED # USED To make things worse, when ink strokes were painted over, sometimes the # USED ink would dissolve and either would spread around, or would be pushed by # USED the brush for a small distance; as seems to have happened at (D). # USED Presumably, some strokes were completely washed away and spread over a # USED larger area, mixed with the green paint. # USED # USED At (E) there are some smudges which look like those at ( C) or at the # USED nymph's toes, and they seem to form a rounded shape that does not seem # USED to be just a random stain or vellum defects. Will we be able to # USED determine if it was indeed painted-over ink, and, if so, recover enough # USED of it to tell what it is? # USED # USED All the best, --stolfi # USED ###################################################################### # USED 21-10-2025, 05:12 PM (21-10-2025, 12:53 PM)Searcher Wrote: Тry on this # USED fragment: # USED # USED Hm, unfortunately there is some bleedthrough of the dark brown ink of # USED the other side (f22r) that may be responsible for those darker streaks # USED on that clip: A B C Image A is a clip of f22v that covers yoru area of # USED interest. Image B is a clip of the matching area on f22r (as accurately # USED as I could determine it), flipped left-to-right for convenient # USED comparison. Image C is image B with inverted colors and ~50% # USED transparency overlaid on image A. # USED # USED On the left half of mage A, notice the faint ghosts of the berries of # USED image B. On image C, notice that there are many berries on the other # USED side of your area of interest. # USED # USED However, the ghosts are not that strong, so maybe they don't account for # USED all the dark smudges in the area of interest. I will see what I can # USED do... # USED # USED All the best, --stolfi # USED ###################################################################### # USED 22-10-2025, 11:58 AM (21-10-2025, 09:21 PM)Cile cile Wrote: What about # USED f35r, left side of the plant? # USED # USED Wait, wait -- I am still working on the first real case... # USED # USED But I don't care much about the Herbal images, since I believe that they # USED are mostly random "Frankenstein monsters" created by the Scribe. Many # USED leaf veins and outlines have been painted over, but I don't think they # USED carry any useful information. So those puzzles will get very low # USED priority in my queue... # USED # USED The figures in Bio, on the other hand, seem to be more meaningful, even # USED if they are disguised -- as organs or as baths, depending on your theory # USED -- and have been badly affected by fading and retracing. Any drawing # USED details that were obscured by those green pools may perhaps give us # USED additional clues about the nature of the drawings. # USED # USED All the best, --stolfi # USED ###################################################################### # USED 22-10-2025, 11:08 PM [RENE] # USED # USED I have been trying off and on over the years # USED to 'detect' ink under the paint, by processing the images. # USED # USED The case without paint is already difficult. This is because the basic # USED hue of the parchment and the ink is relatively close. This is also # USED visible from the color cubes shown in one of the earlier posts, as the # USED two lines are roughly each other's extensions. All one has to go by is # USED the lightness. # USED # USED When there is (say) green paint, at least we have a hue difference, but # USED dark paint is still very hard to distinguish from paint over ink. At # USED least using standard approaches. AI-based approaches should have a # USED better chance here. # USED # USED However, in all cases I had to conclude that the human eye and brain # USED together are the best tool to recognise things. Using image processing # USED to enhance the image in whatever way, rarely ever brought anything that # USED wasn't already somehow visible and used by the brain. (But at the same # USED time lots of stuff that was meaningless). # USED ###################################################################### # USED 23-10-2025, 02:43 AM (22-10-2025, 07:58 PM)Cile cile Wrote: # USED I meant about writing under the dye Smile # USED # USED # USED OK, let's have a look... # USED # USED Hm, there does seem to be writing inside that big ocher-painted area, # USED but is it just bleedthrough or offset? # USED # USED Offset from f34v can be excluded since there is no writing on f34v # USED anywhere near that area. It could be offset from some other page, that # USED happened back when the bifolios were still unbound. # USED # USED But it does look like bleedthrough from f35v. Here is a clip of f35v # USED corresponding to that area, flipped L-R for easy comparison: # USED # USED It does look like most if not all of the "writing under the ocher paint" # USED is indeed bleed-through of the writing on f35v. Here is the second clip # USED with colors inverted and transparent background, overlaid on the first # USED clip: # USED # USED Do you see any bits of writing or drawing that is not bleed-through? If # USED you do, please point it out to me and I could try to extract it... # USED # USED All the best, --stolfi # USED ###################################################################### # USED 23-10-2025, 09:21 AM [KOEN] There appear to be various somewhat regular # USED curved shapes to the right of your blue letters. They almost seem to # USED follow the edge of the drawing, like decoration. My blog: The Voynich # USED Temple # USED ###################################################################### # USED 23-10-2025, 10:13 [OSHFDK] (23-10-2025, 09:21 AM)Koen G Wrote: There # USED appear to be various somewhat regular curved shapes to the right of your # USED blue letters. They almost seem to follow the edge of the drawing, like # USED decoration. # USED # USED These are visible on my decoloured Voynich images too: # USED https://oshfdkbw.pages.dev/images/image.069.jpg # USED # USED Look like small circles all over the bulb. # USED