I've been toying with ways to "depaint" the manuscript. The most recent work I can find is Landini's colour deconvolution from back in the early 2000s. I wanted to check if anyone knew of newer developments? Also, is there a significant set of "depainted" images anywhere? I can only find fragments of Landini's image results, and some cruder image work done by Pelling (focused on recovering painted-over text).
Thanks!
Eric
There is a vaguely related thread here : The Voynich Ninja > Voynich Research > Imagery >Image processing, highlighting hidden details
You are not allowed to view links.
Register or
Login to view.
That was a bit of a spin-off of other image processing I have been doing.
I have actually been very interested in colour removal, and got to a reasonable level of success, but not nearly as 'reasonable' as I would have hoped.
One user even made an animation out of it.
Let me see if I can find the thread: You are not allowed to view links.
Register or
Login to view.
Edit: let me see if I can fix the broken links.
I imagine that to get really good results of line art as it must have looked, some tedious manual cleaning up (and thus interpreting) would be required. Unless this could be done by some AI process.
I've been working on a machine learning based solution that would separate the image into layers of "vellum/ink/dirt (stains that are hard to classify)/scratches and defects/red paint/green paint/etc...", based on sampling of 3x3 pixel areas of the visible light TIFFs. The results so far have not been as good as I expected, it would detect traces of ink and paint in many random places, not as some distinct shapes, but more like blobs of random noise. And if I set a high enough detection threshold, then quite visible portions of ink or paint won't be classified correctly. At one point I realized, that there could very well be ink/paint traces all over the place due to ink/paint/tinted vellum particles shedding from opposite pages. So it's possible that there is no solution that can reliably separate layers using only local pixel values.
One can train a model that would have spacial awareness (working on colors AND shapes) and it might fare much better, but I'm personally against this, since models with spacial awareness will most likely introduce in their results some shapes that are not present in the original MS, and this will make any analysis based on the results from these models hard to validate.
(05-12-2024, 08:25 AM)Koen G Wrote: You are not allowed to view links. Register or Login to view.I imagine that to get really good results of line art as it must have looked, some tedious manual cleaning up (and thus interpreting) would be required. Unless this could be done by some AI process.
The old pictures did involve quite some manual editing. In the mean time have Linux at home, and can achieve the same fully automatically with home-brewn software. It is far from perfect still.
AI is certainly the right tool for this type of task, but I prefer to do it myself. More fun that way.
(05-12-2024, 10:56 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.The old pictures did involve quite some manual editing. In the mean time have Linux at home, and can achieve the same fully automatically with home-brewn software. It is far from perfect still.
AI is certainly the right tool for this type of task, but I prefer to do it myself. More fun that way.
Hi Rene,
I also have a Linux system. Can you perhaps offer your software and introduce the function ?
One day I could do that, but now it is not practically possible.
This is not just because it is all in fortran (77) but mainly because it is part of a larger integrated system that revolves around a database. Rather: a database-like structure of files.
The principle behind this application is that for every pixel in the image I define a 'mask' with 8 bits.
These bits encode whether the pixel has: ink (writing vs. drawing), paint, both or neither, or whether it is outside of the net page. Not all of that can be automatically detected. The file of mask bits is itself a BMP file, and all of my software reads and writes BMP files only. These are bulky but their access is straightforward.
There is one set of tools to set up the mask file (always 'under work'), and then another tool to use the mask with the image to create a 'derived image'. That second tool is the easy part.
(04-12-2024, 06:50 PM)evandrunen Wrote: You are not allowed to view links. Register or Login to view.I can only find fragments of Landini's image results
Gabriel Landini's plugin for ImageJ is free so here is a test result on the green color:
[
attachment=9497]
Substracted:
[
attachment=9495]
(06-12-2024, 10:11 AM)nablator Wrote: You are not allowed to view links. Register or Login to view. (04-12-2024, 06:50 PM)evandrunen Wrote: You are not allowed to view links. Register or Login to view.I can only find fragments of Landini's image results
Gabriel Landini's plugin for ImageJ is free so here is a test result on the green color:
Substracted:
Thanks. I'm mostly using GIMP and Python, and have implemented a method similar to Landini's as a plug-in. I was just wondering if somebody had already run it on the whole manuscript and had it available somewhere so I could browse through it and not duplicate the effort. Spot results start with "amazing!" and then devolve to "not as good as I'd hoped" (generally, the green paint is the easiest to deal with), which applies to all the different approaches I've taken and seen others attempt so far. It's a fun puzzle to chip away at.