John Leimgruber:
tl;dr;
i made a denoise/deconvolution-ish tool https://ubersmooth.com to help with this in my images (free download for windows, put in 0 and click the paypal button [no need for any accounts]). it is bare bones, command line interface, no graphical mouse stuff.
background
I used to use the Richardson-Lucy Deconvolution in siril, early in the linear workflow stage. I'd basically crank it up as much as possible before big rings around the stars, then back it off a little and call it good.
It does sharpen everything up a bit, reduce star size a bit (not so good at fixing mis-shapen stars), but increases high frequency noise as well.
I realized the PI folks were all using RC Astro BlurXterminator and looked into it a bit. I tried astrosharp and some others early on, but didn't like the strange artifacts.
After reading some papers on U-Net and finding an example model, i fooled around with python's pytorch and got something working okay. To train it i used nasa images then added gaussian blur and added random mask noise to bloat stars. I would present the "original clean" version and the "artifical noise" version to the model for training on my old GTX 1070 GPU. Took a while, but it did start improving some images.
It's not perfect by any means. To improve it I'd need to get a bigger data set and improve the star treatment by using actual plate solving on every image as currently it confuses Hii regions in small galaxies for stars sometimes etc. There are a couple few ai models to try, the "planetary" one was trained without any star treatment.
ymmv but it is another tool for any experienced command line tool users.
as ubersmooth is run *last* in the process at the very end, i combine it with GraXperts denoise and seems to work well. Maybe GraXpert will build in a BlurX style star treatment / deconvolution model eventually...
cheers and clear skies!
-john (ubergarm)
Good luck with ubersmooth. I suspect I read some very similar papers and intended to do something like that myself, but never got round to it. The more options that are available, the better!
