(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.