Discussion about the current problems with AI-assisted star removal with StarX/Starnet [Deep Sky] Processing techniques · Jackson Datkuliak · ... · 7 · 367 · 4

FIRST I want to say that I am NOT trying to self promote here! My video is UNMONETIZED and I am looking for genuine discussion on how we as a community can improve the star removal techniques we currently employ in processing.

With that being said, I made a video talking about an M31 image I am processing from Starfront Remote Observatories that I am having a lot of trouble with star removal on. My problem is that there are 10's of thousands of galaxies in the background of this image (for real, I checked with SDSS!), so it is effectively impossible to have a good star removal. Star removal on good M31 images already sucks because they tend to take a lot of bright nebulosity and globs within m31, but they also tend to take thousands of background galaxies too. You don't need to watch my video to contribute to the discussion, but in my video I talk about what methods I have tried (and failed) to use to get a perfect milky way star removal (removing only stars within our milky way from the image).

If you would like to watch the video, this is the link https://www.youtube.com/watch?v=eDLqA3bRBmM

And if you would like to see 10's of thousands of background galaxies (can literally see entire galaxy clusters), check out this image on AB: https://www.astrobin.com/qfhk7a/C/?nc=collection&nce=39541 I would also recommend if you are looking at the M31 image to download the original, full resolution tif file thats linked in the description of the post!

Also, all of the data I use and am talking about is free for anyone to download. Its hosted at https://data.bortle.org/tag/46/starfront-remote-observatory and there are currently a 30 hour LRGB M31 (the dataset I'm talking about in this post), a 40 hour SHO NAN, and a 20 hour LRGB iris and ghost nebulae! All of these projects are still a WIP, but you can download what we have collected so far and give it a process for yourself!

andrea tasselli:
As a matter of principle I never watch YT so I'll skip the peculiarities of the specific sample but in general SN does a much better job at removing stars in the linear phase than SX does. This said, even SN fails at time so in both cases the AI isn't up there where it would need to be. Bottom line: you need to learn to process the whole image a single unit, as it was done before SN came to the fore.

yeah if you watched the video you would know that it’s more about developing new, different processing techniques. i have no problem making a fantastic image without star removal, but my idea is to be able to remove the stars so you can see the structure of galaxy clusters and filaments of galaxies in the background at a large scale. can’t do that with stars in the image as the stars will dominate everything even if they are reduced/understretched

i also talk about how i think new methods for better star removal could be really interesting for better science communication and display of the images we take. if i can actually remove JUST the stars from my m31 image, you could see filaments of galaxies in the background throughout the whole image. it’d be game changing, there would be no other image like it. just a thought experiment i’ve been throwing around

andrea tasselli:
Well, I can't see 10^4 background galaxies in that image but so far I can't recall having issues showing small background galaxies and galaxies clusters either with or without star removal. Obviously it helps having an optical systems with relatively small PSF and large image scale, IOW a fast large telescope.

Again, it's not about that I can't show them. A simple LRGB combination and autostretch shows everything. It's just impossible for a human to see the small galaxies when looking at the entire image. I want to be able to look at the entire image and be able to see the filaments of galaxies in the background ideally. 

But either way, there are really thousands of galaxies in the background. Here are some crops showing a couple of the galaxy clusters visible. Fun stuff to try and find looking around



I think this also helps highlight how bright the stars are in comparison. Just not something the human eye is gonna make out with zooming way in.

Also, running SDSS on the image with galaxies only returned >20,000 hits. About 18,000 up to mag 20 iirc, and looking through the image I couldn't find any on the up to mag 20 image that I couldn't make out from the sky background.

Wei-Hao Wang:
Can't you just use the SDSS data to mask the galaxies to prevent them from being removed?  Is this cheating?  If using Gaia data to help us to calibrate the color of our images is not cheating, using SDSS data for star-galaxy separation definitely isn't cheating.

Yes you can! And I tired. Unfortunately, SDSS only covers part of the sky, so it won't work for every image. It covers almost all of my M31 image, but the problem is more that actually making the mask is a nearly impossible task. There tens of thousands of galaxies that SDSS has catalouged in the image, so many that I can't actually annotate all of the ones that are visible in the image with the 20,000 object limitation of the AnnotateImage script in PixInsight. And yes, this is also with the limiting magnitude filter. I tried sitting in photoshop and manually masking out all of the galaxies but it was really futile. After over 2 hours I had barely made a dent in the image. If there was some way to make a script that automatically detected PSFs and then referenced their exact RA/DEC from SDSS to determine if it was a star or galaxy, that would be great. However making the mask still would have challenges as there are lots of galaxies in this image that lie behind a bright stars halo. It really is a difficult problem lol. Manually masking 20,000+ galaxies with the Photoshop paint brush tool doesn't seem like a good final solution as it would take tens of hours of work for each image, again not to mention that SDSS doesn't cover the whole sky.