I've also tried applying the same model to the top of 116v, without retraining, to see how well it would generalize detecting ink. It worked without any problems on the first try, I just had to make sure the set of files is the same. Images for 116v include a separate set of "TX" images that appear to be taken using backlighting. There are no such images for 1r.
Overall, there is no magic, the model just makes the text a bit clearer by combining the information from all the MSI images. Maybe next weekend I'll try training a separate model on 116v specifically to see how much difference it will make (there are a few obvious places here where faint ink is not detected with the 1r model).
Here's what the model trained on the margin of 1r produces for the top part of 116v:
Full resolution image: You are not allowed to view links. Register or Login to view.
And this is an example of a TX image of 116v from the MSI collection (no processing other than applying gamma to squeeze 16 bit image into 8 bit):
Overall, there is no magic, the model just makes the text a bit clearer by combining the information from all the MSI images. Maybe next weekend I'll try training a separate model on 116v specifically to see how much difference it will make (there are a few obvious places here where faint ink is not detected with the 1r model).
Here's what the model trained on the margin of 1r produces for the top part of 116v:
Full resolution image: You are not allowed to view links. Register or Login to view.
And this is an example of a TX image of 116v from the MSI collection (no processing other than applying gamma to squeeze 16 bit image into 8 bit):
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)