Why I'm using DeepSNR For Noise Removal and You Can Too

SMALL CORRECTION - my stretched image is NON-LINEAR .. not linear like I say in the video and I shouldn't have left the deepSNR box ticked... thanks for calling me out on my error! The speckle however likely doesn't come from this but from DeepSNR being trained on drizzled noise scales which @Ekuy1 notes in the comments. Scaling down matched the noise scale so the model was better able to denoise smoothly - interesting!

Loving the more regular content mate, keep em coming

Beautiful image, Dylan! The nebulae you capture in there with HaGB are stunning!

dsnr is trained on properly drizzled data, and doesn't work well on standard integrations. this is because proper drizzle has scale 1 noise, where each pixel's noise profile is independent of the surrounding pixels (its more complicated but go with this for now), while standard integrations use various interpolation methods which will cause larger scale noise profiles, such as 2 or 4 pixel noise. due to the nature of dsnr, these clusters in the noise will be mistaken for stars, which will cause the "speckling" you see here. by resampling data down by 2x which has improperly interpolated, 2-pixel noise profiles, like you've done here, you essentially reduce the noise scale to single pixels, which is what dsnr has been trained on, and recognises as well pre-processed data. if you want your images at full resolution, and not scaled down, use WBPP drizzle, or another algorithm which correctly implements drizzle integration (Siril and DSS drizzle do not work, they still use surrounding-dependent interpolation), then dsnr will work perfectly well on full-resolution images, just like the results you're currently getting right now. also, drizzle definitely DOES work on CFA (or debayered) images, but likewise, it needs to be correctly drizzled. p.s. lots of misinformation and misconceptions around the drizzle and deepsnr algorithms on cloudynights, if you want to learn more, read the source documentation from the authors "Variable-Pixel Linear Reconstruction"

Great vid, as always! Looks like another good plug-in for my PixInsight. Thanks, buddy.

Thank you for teaching me over the last 4 yrs I have been loving this astro stuff! There are a couple others as well but wanted to say Thanks!

Goin' deep again! woohoo!

Thx for the educational video Dylan, from Argentina, another southern Chad sky

Down sampling indeed makes everything better. 🎉

Awesome info man.

I just tried this on a non-linear image and it worked beautifully! I had Starnet set up for GPU acceleration and this used the GPU as well with just a normal installation. Great software!

Dylan, with pixinsight's WBPP stacking menu, use these settings- Drizzle scale 1x Drop Shrink 0.85 Varshape = 1.5 Grid Size is variable based on dither distance. I dither 30px at 0.58"/Px, so I use Grid Size of 8. if you're under that, consider 16, or even 24 or 32. Increased Dither with Drizzle and DeepSNR will give you a massive boost in SNR and detail, I've measured an increase of ~5-10% in both sharpness and SNR, plus you get the ultra clean drizzle background.

Thanks Dylan.

DeepSNR is a great tool. I’m just eagerly waiting for gray scale support because I don’t want to have to run 3 separate integrations and run it on an rgb combination of that.

if you CFA drizzle your data at 1x then the dots disappear when using DeepSNR so you can CFA drizzle your 10000 pixel-wide image at 1x and keep it at 10000 pixels wide and still get the super clean result :)

Great video, instructive and clear! I was wondering if this software can also be installed in Siril...?

Thanks Dylan will check this out. Can you do a video on resampling and sizes as you mention here I am not doing this with drizzled data and you are making me think I should be? Thanks

I never liked Topaz noise reduction. The biggest problem to be fair was probably my own inexperience and the fact I was running it on an older computer and OS which although according to Topaz was compatible, was not really up to it. When I upgraded to a new and much more powerful laptop with Windows 11 I started using Noise Exterminator which I find gives much better results. I bought this as a package with the star and gradient exterminator plugins which also work very well, I would have liked to have bought blur exterminator as well but as I don't use Pixinsight it wouldn't work. Nice image of M83, DeepSNR does look to give great results.

This is cool, I definitely need to try DeepSNR and downscaling earlier in the process. Disclaimer: I have NO idea what I'm doing most of the time lol. I usually start out with a bit of BlurX followed my NoiseX as the first two steps in processing. I'm nowhere near the level of astrophotographer as you are, but it seems like a good thing to do before using StarX or doing any stretching. Lately my processing has involved a lot of experimentation with the order of things I do to see what works best but I can't really land on a solid permanent order. After some basic stretching on both stars and background independently, I'll take things into photoshop where I'm a bit more familiar to finalize as separate layers. One of my last steps is usually hitting the background layer with another bit of NoiseX or Topaz and I'll play around with them both to see what I like better. I don't normally downscale until I'm exporting a final I’mage because I like the option of having a full resolution photo in case I need it later while being able to export smaller sizes for social media at the same time. Sometimes I wish there was an easy was to process every image the same way, but I guess if it was easy then everyone would do it.

Wow !! That software is so good !! Why do we even need telescopes ANYMORE !! Just take a picture through your binoculars with your IPhone and PRESS THE HUBBLE BUTTON in our AI software !! LOVING IT !! 😅