Upsampling vs Drizzling
ฝัง
- เผยแพร่เมื่อ 5 พ.ค. 2024
- If you're developing information from a telescope and camera that were perfectly sampled for the seeing conditions, wouldn't it be better to upsample than drizzle? I was recently posed this question by a viewer, and it's a good question. In this video, we'll process a tough DSO, the Ghost Nebula, with well sampled information and see whether upsampling or drizzling produces a better result.
Too the full res outcomes, follow this link: www.astrobin.com/wzvztq/E/
To see the final image of the Ghost Nebula, visit the Sky Story Astrobin gallery at: www.astrobin.com/xorr7d/
*** *** ***
Check out a portfolio of astro imagery shot from the Sky Story Observatory at Astrobin: www.astrobin.com/users/SkyStory/
We are Telescope Canada's first affiliate. It's a great place for astrophotography gear. Using the following link adds nothing to your cost but helps us produce more educational videos about this amazing universe around us. telescopescanada.ca?bg_ref=KK15sLXmWw
All videos best viewed in 4K.
Peer into the cosmos on Sky Story and along the way come to appreciate our world all the more. Here, we explore the realm of nature beyond our Earth through the fields of astronomy and astrophotography.
All our programs are committed to the standards of scientifically accurate, high quality content that aims to educate and inspire. New videos are posted almost weekly, so please subscribe.
#astronomy #astrophotography #space #nebula #stars
Great video👍
I'll take your word for it cos I honestly can't see much difference on TH-cam.
Yeah, YT has terrible compression. The differences here are subtle, so it helps if you can view it in 4K, and it's even better if your monitor is color calibrated.
@@SKYST0RY I was going to say the same thing. I switched to 4K and it's still really hard to tell the differences via youtube. It probably makes sense if you actually post links to the raw files so people can do the same comparisons you are doing. Otherwise, it seems we're just taking your word for it. It is also hard to tell if some of this is slight differences in curves/stretching as well. Because even really tiny differences to a stretch/curves can show a significant difference in the results in my experience.
I will still always drizzle based on your suggestions, I don't disagree with the base idea here. I would also be curious to know if you see similar drizzle benefits when not using BXT in your workflow. E.g. how much of it is truly from Drizzle or from the algorithms in BXT and how they are affected by drizzling.
I am sorry, I forgot to link it. I have placed full res versions of all the outcomes on the Sky Story Astrobin that you can inspect. www.astrobin.com/wzvztq/E/
I am sorry, I forgot to link it. I have placed full res versions of all the outcomes on the Sky Story Astrobin that you can inspect. www.astrobin.com/wzvztq/E/
Thanks for this, very interesting and I have found drizzling on some of my data which is very oversampled seems to make noticeable improvements to noise and fine detail.
It's quite funny but I actually preferred the unprocessed image on an overall level on a tablet but I don't think it's really possible to make a proper evaluation on youtube compressed videos without a decent monitor.
I would also like to compare the images without the use of AI tools, it seems hard to know how much they are disproportionately influencing the outcome.
Also very few people can capture data of this quality, it would be interesting to run these comparisons on poorer quality data with a level of correlated noise etc.
These are good observations. It is absolutely true that TH-cam is a nightmare to work with due to video compression. Its compression is slightly logarithmic, too, compressing darks harder, which is really bad for black crushing in so many astrophotos unless one artificially and heavily lifts the blacks. (Probably gives rise to the myth some astrophotographers have that space is not black or should be portrayed as a pale gray.)
Persons have asked in the past if the images would still come out better without AI tools. I have run processing test sequences without such tools and drizzling still exceeds the quality of non-drizzled images. It is important to understand that all AI is is an algorithm. It selectively runs something like an unsharp operation over an image, and the "training" done with AI is just teaching the algorithm on what to run it. This is much the same as how noise tools work by applying formulae to attempt to determine what is background vs noise. The AI isn't sentient, it doesn't really "learn". It's just a flexible algorithm, a kind of better implemented old fuzzy logic. What this means is the line between what is accomplished by an AI tool vs a non-AI tool is pretty blurry. At their core, they're all just algorithms. Even when I run BXT, I still choose the power and values it applies to an image, and when in the developing sequence to run it. Otherwise it does a poor job. Very much like any other developing tool.
I have stopped running such tests, though. I have come to feel such tests are arbitrarily crippling image development to see if an outcome is "valid" without a tool. I'd rather apply every tool at my disposal to bring the best out of the information. Validity is determined by the nature of the experiment. So long as the tool isn't simply making stuff up in the manner of CGI, the valid outcome I am looking for is the best possible image.
Nice music back ground whats it called? Great video
The music is called Azara.
I appreciate you doing these. I resample 50% my 2x drizzled images for the sole reason of better SNR according to my eyes and the PI SNR script. I have not compared in detail to determine if there are any differences in detail. I pixel peep and watching the evolution of some of these stars in your video was interesting especially the medium brightness light orange ones. There was one in particular that sits just above the curve of the nebula to the right. It looks like a double star with a red companion. In the non-drizzled, non resampled image it looks like a double star with just a hint of a black artifact above it. In fact, I can see a hint of a black artifact just above several of these medium brightness, orangish stars. With the 2x resample much of the small red star is mostly lost in the brighter star's glow and the black artifact above the star is enhanced. With the 2x drizzle, the two stars are now clearly separated and the black artifact is even more enhanced. When I go to your final image in AB, there are several stars that the black artifact above them is clearly visible when cropped in, but that red companion of the star I was talking about is completely gone. Looking at Stellarium, that small red star exists and can be seen on the YT video but not in the final image.
What do you think is causing the black artifact above some of the stars? BXT?
What would have happened in the final processing that would have wiped out that red companion star?
No drizzle is the best for what i see.
I have experimented with that in the past. The artifacts result from SXT. It seems to do better extracting stars from either non-drizzled or drizzled 1x images, and drizzling 2x doesn't exactly place nice with stars. I think it was in another video I had recommended drizzle 1x for stars and 2x for non-stellar images. Some of the star artifacts, however, are the result of the SCT's dew ring heater affecting the corrector plate, a known issue with the dew ring heaters. I believe I will have resolved within a week by going about a new protocol to mitigate dew.
Keep testing and sharing my friend, let’s find the truth… thanks for the effort.
this a lesson in critical image evaluation. Not gonna lie, it took some effort to follow the subtle changes.
TH-cam compression causes some loss of detail, but the differences are subtle.
I’ve always heard to drizzle at least 1x on mono data. Any thoughts on that?
I just ran a test of drizzling 1x on the LRGB data I used to make this video. 1x drizzling produced better stars than the 2x drizzling but 2x drizzling produced better DSO structure.
@@SKYST0RY I was going to ask this same question. Just need to get WBPP to allow two different drizzles in the same run so we can use the stars from one and the DSO from the other!
Given you are showing the 2x drizzle at 1:2 when comparing with the original, which is the equivalent of an IntegerResample of 2 on the image - is it best to do that at the end and get back to the original resolution, or is the 2x resolution change something you want to keep at the end?
@@zaphus I'm not sure I entirely follow. Processing and drizzling x2 and x2 were done with WBPP before any further processing. I always keep the 2x drizzle when I am done processing them.
@@SKYST0RY Sorry, I both didn't explain well, and got the facts wrong :-)
Initially in the video you are showing the 2x images next to the 1x, with the 2x downsampled so they look the same, but I see now that later you are doing the opposite, viewing the original at 2:1 and the 2x at 1:1.
My question was really around whether you see any benefit doing an integer resample to bring the 2x back to the original resolution or not, which you have answered (you keep it at 2x).
@@zaphus On Cuiv, the Lazy Geek's channel, he experimented with down sampling after drizzling and found it helped suppress noise. I experimented with the technique, though, and found the cost of the process was not worth the benefit, at least not with my workflow.
Interesting video.
What about drizzle 1x? That might actually give the best result.
Cuiv, the Lazy Geek tested this on his channel and found it resulted in improved noise management but since NXT I don't find it necessary. But you have me curious, so I am presently restacking the data with a 1x drizzle. We'll see how it goes.
I just ran a test of drizzling 1x on the LRGB data I used to make this video. 1x drizzling produced better stars than the 2x drizzling but 2x drizzling produced better DSO structure.
@@SKYST0RYMaybe 2x for structure and 1x with StarX for stars and recombine for the best overall results?
likely so. drizzle 1x should have the best result overall, with the true non-interpolated noise profile not-present in standard integrations, while having better SNR than drizzle 2x, as its signal:noise is being distributed into 4x less "drop wells" than 2x, theoretically having a 4x better SNR assuming perfect gaussian noise and drop size 0 (interlacing basically). in theory, the noise would be neither perfectly random nor would you use a drop size 0, so the benefit would be less defined
@@Ekuy1 I wonder if 2x drizzle is a sweet spot for BXT, and this is why @SKYSTORY sees better results in DSO structures... rather than it being primarily a benefit of drizzle x2 itself. Additionally, some comparisons between BXT and Hubble have previously shown that BXT can sometimes invent faint structures. So another level of interesting testing would be to see if drizzle x2 + BXT just gives nicer looking images, but further diverges from real data as seen by hubble. Maybe some good ideas for more videos. Either way, great thought inducing content from SKYSTORY regardless.
Okay.
Is it just you tube or do the red stars in drizzle look like hot pixels? Otherwise drizzle is better.. but those "re dot" pixel stars don't look good, atleast on you tube
You're absolutely correct. The red stars don't look great. It's not TH-cam's fault. I wanted to do as little developing of the image as possible to keep the comparison apples to apples as much as possible, so I didn't color balance the RGB information. I also skipped about a dozen other steps. To see a final drizzled x2 image, follow this link: www.astrobin.com/xorr7d/