Seeking NoiseXTerminator and GraXpert Denoise Preferences and Recommendations in PixInsight - Forums

Greetings!

I've been comparing denoising results in PixInsight using various settings of the GraXpert Denoise script and NoiseXTerminator process. I would appreciate hearing how others are using either one or both of these excellent tools. In the following I provide some results that, at least to me, are essentially identical, having adjusted the respective parameters to my taste. I am sure, however, that each method emphasizes a different process and set of decisions, and it is the difference of these details that interests me.

The following are zoomed-in cropped images of NGC 5906, The Splinter Galaxy. This is the result of about two hours integration time with a 10" TS RC, QHY600M-PH, TSRCFlat3, FL 2000mm, resolution 0.38 "/px. The processing up to this point has included ImageIntegration, background subtraction using GraXpert, then BlurXTerminator. Following that I applied to separate instances of the image (1) GraXpert Denoise and (2) NoiseXTerminator, for comparison. In both cases, I applied the default stretch using ScreenTransformFunction applied to HistogramTransform. The denoising input settings and Statistics outputs follow.

The default settings of NoiseXTerminator produce a more aggressive noise reduction than I prefer, whereas GraphXpert Denoise's default produces a milder reduction for my taste, therefore I played around with both until I got something that looks and measures nearly the same, at least according to the avgDev and MAD values in the following:

This is about as close as I could get for the statistics, and, to my eye, the two images are nearly identical with only the small subtle differences perhaps in star halo. This suggests to me that I could easily choose one or the other. However, I don't want to potentially miss out on some details that may turn out to be important to me.

Again, I'd appreciate your input and recommendations on how to best use one or both of these tools, explain perhaps what you like about one over the other, provide details on how they differ, etc.

Thank you, and I look forward to hearing your responses.

Best Regards,
Ben

how I use it?

well, with NXT, first of all, I never ever increase the Detail slider, because IMO, all it does is damage details, so I leave it at 0 always

so, for a very noisy image, usually it's 70-80 for Denoise

for the majority of the datasets (the ones that are neither too noisy nor too clean), it's usually 50-60 for Denoise

and for the cleanest data, it's usually 20-40

not sure if that's what you wanted to hear; I'm not a very statistical person, like some of the other common chatterboxes on AB

so, sorry if I couldn't explain more why I choose those settings, but those settings work for me always

and I didn't like Graxpert the first time I tried it for denoising; seemed to me it still didn't compare to NXT, but maybe I used it wrong; I never use Graxpert for denoising anymore.

how I use it?

well, with NXT, first of all, I never ever increase the Detail slider, because IMO, all it does is damage details, so I leave it at 0 always

so, for a very noisy image, usually it's 70-80 for Denoise

for the majority of the datasets (the ones that are neither too noisy nor too clean), it's usually 50-60 for Denoise

and for the cleanest data, it's usually 20-40

not sure if that's what you wanted to hear; I'm not a very statistical person, like some of the other common chatterboxes on AB

so, sorry if I couldn't explain more why I choose those settings, but those settings work for me always

and I didn't like Graxpert the first time I tried it for denoising; seemed to me it still didn't compare to NXT, but maybe I used it wrong; I never use Graxpert for denoising anymore.

I don't know anyone who sets the detail slide to 0... Unless you have a really good reason, you should use the defaults of NXT which are 90/15.
I would think Russell Croman sets the defaults for a reason.
I went to a local Pixinsight class taught by Kevin Moorfield and he uses 75/15 for stuff he shoots in Chile.

Dan Kearl:
how I use it?

well, with NXT, first of all, I never ever increase the Detail slider, because IMO, all it does is damage details, so I leave it at 0 always

so, for a very noisy image, usually it's 70-80 for Denoise

for the majority of the datasets (the ones that are neither too noisy nor too clean), it's usually 50-60 for Denoise

and for the cleanest data, it's usually 20-40

not sure if that's what you wanted to hear; I'm not a very statistical person, like some of the other common chatterboxes on AB

so, sorry if I couldn't explain more why I choose those settings, but those settings work for me always

and I didn't like Graxpert the first time I tried it for denoising; seemed to me it still didn't compare to NXT, but maybe I used it wrong; I never use Graxpert for denoising anymore.

I don't know anyone who sets the detail slide to 0... Unless you have a really good reason, you should use the defaults of NXT which are 90/15.
I would think Russell Croman sets the defaults for a reason.
I went to a local Pixinsight class taught by Kevin Moorfield and he uses 75/15 for stuff he shoots in Chile.

Well now you know someone

last time I remember using Detail, I saw it made a sharpening effect similar to the cheap sharpening tools in Photoshop, which is garbage

but I'll try it out again and post here the difference between using Detail at 15 and Detail at 0; I'll probably use a galaxy for the comparison

Well now you know someone

Thanks for the snark... suit yourself, if you think you know better than the person who wrote and designed the software, Good for you.

I thought you literally meant that you don't know anyone who sets the slider to 0

If I said the truth, why are you offended?

I thought you literally meant that you don't know anyone who sets the slider to 0

Yes, I know no one who sets the detail in NXT to 0. Why do you?
If you think Russ Croman's tools are like "cheap" tools in photoshop, don't use them.
I think most Astro people appreciate his software and most use the default settings or close to them.
If you know better, fine, What is your point?
Just do your own deconvolution the manual way.

I also set detail in NXT to 0 for the same reason. However, nowadays, I rarely use NXT. DeepSNR is much more powerful if used carefully.

Maybe it would be a good idea to revisit exactly what NXT is doing with respect to PSF and the best ways to implement it. Here is an excellent video by Adam Block…..https://youtu.be/6hkVBnYYlss?si=3XPoyhcQtIcovw1M

From my experience NoiseXterminator did a great job of eliminating noise - but unfortunately created easily artefacts that create heterogenous blobs during stretching.

So I played around with GraxPert denoise - and I am blown away, as I get a better noise reduction with zero risk of artefacts, thus I have altered my process to only user Graxpert Denoise.

Arny

IMHO The order in which these tools are used is critical and the values to be used are determined by the nature of the field and its objects.

I use previews to test the settings that work best at the time. NoiseXTerminator is only used much, much latter. There was some advice a while back regarding how the AI was trained to support that. I know some like to use it early BUT BlurXTerminator does NOT like de-noised data.

My current workflow is GraxPert background subtraction FIRST on the un-combined stacked images generated from WBP, typically with Maximum Smoothness.

The new GradientCorrection script is a work in progress and can produce great results if you play around with it. When MAP is released THAT should be the bees knees hands down winner by miles.

I found Graxpert noise reduction to pretty slow (that's being generous ... zzzz .. wake me up when we get there) compared to NoiseXTerminator but it's free so no grounds for complaint. Maybe the latest version and the latest AI library are quicker.

Then BlurXTerminator Correction ONLY applied to each channel before finally combining.

Once the channels are combined and Spectro Photometric Color calibration has been applied, and Narrow Band Normalisation for SHO, I use BlurXTerminator for Star Reduction and some Non-Stellar Adjustment.

I then generate Starless and Stars from the Linear state so they don't mess with each other.

I typically stretch the Starless with GHS, setting the Linear black point first. Initial soft stretch in Color Mode is my preference. There are plenty of alternatives for stretching but GHS seems to provide the best control.

For stretching the stars it's a simple Histogram Transformation boost (and for SHO once in Non-Linear, Invert, SCNR for the green and Invert back) before Curves Transformation to boost the star's color saturation (testing the results with Pixel Math as I go). I also use an ACDNR Lightness mask once I am out of Linear to separate the Stars and background and deal with each as desired. That also applies to Starless. ACDNR Lightness masks are pure gold.

Again use a preview and check the results step by step.

I find actively using the NoiseXTerminator detail setting works great when used late but it's a loaded gun so test, test, test before pulling the trigger.

Good Luck.

Ian

@Ian Parr , very interesting post. One question if I may--what is MAP?

There is religion and there is science. And they don't mix well…

Why would Russel have decided on a detail slider setting of 15 but given the user at the same time the will to slide it all the way down to zero or up to 100? Is there some sort of sacrilegious astronomical sin committed by touching that slider? LOL

I have done several tests with both tools. Although they achieve similar results after moving sliders, I personally prefer NXT, maybe because it was the first. However I practically never use it at the default setting. It really depends on the image, if mono or OSC and it's noise level. If I use NXT in linear form, always after star removal. On NB HA around 50-55/10-20. With SII and OIII somewhere in between 60-75/15-25. For OSC linera images I typically use 60-65/10-15. In post-linear I use it at 50-60/10-20. It all depends on your data you are working with and, at the end, you need to like your final image - CS

I don´t know Graxpert. I use only NxT with the following settings:
I do background fixing first, than sharpening with BxT usually in 3 steps. As someone mentioned here BxT should not be applied after Nxt. This gives bad results! I apply for the use with my both Nikon cameras (I dont use yet Astrocams) first "correct only" than I do color calibration and than the first round with setting of ar. 60% and a second round with 40%. The sharpening I set at 5%. Higher values makes the stars look too artificial at least in my configuration.
CS Rudi

Thanks @Ian Parr for listing your workflow…very generous of you to take the time.

The Multiscale All-Sky Reference Survey Project

https://pixinsight.com/mars/index.html

https://youtu.be/ny_0pFOkuTk?si=lOhCSTakK0FQuaIo

https://youtu.be/ny_0pFOkuTk?si=prbzEl9D2So92i1N

They are gathering data now so hopefully soon.

👍

Arny:
From my experience NoiseXterminator did a great job of eliminating noise - but unfortunately created easily artefacts that create heterogenous blobs during stretching.

So I played around with GraxPert denoise - and I am blown away, as I get a better noise reduction with zero risk of artefacts, thus I have altered my process to only user Graxpert Denoise.

Arny

Very interesting, Arny. I'll investigate if I notice this issue. Thank you!

Ian Parr:
IMHO The order in which these tools are used is critical and the values to be used are determined by the nature of the field and its objects.

...

I find actively using the NoiseXTerminator detail setting works great when used late but it's a loaded gun so test, test, test before pulling the trigger.

Good Luck.

Ian

Well said, Ian, and I also thank you for detailing your workflow. Looks similar to the one I've adopted.

Riedl Rudolf:
I don´t know Graxpert. I use only NxT with the following settings:
I do background fixing first, than sharpening with BxT usually in 3 steps. As someone mentioned here BxT should not be applied after Nxt. This gives bad results! I apply for the use with my both Nikon cameras (I dont use yet Astrocams) first "correct only" than I do color calibration and than the first round with setting of ar. 60% and a second round with 40%. The sharpening I set at 5%. Higher values makes the stars look too artificial at least in my configuration.
CS Rudi

In full agreement with you (and others), Rudi, not to apply BlurXTerminator after NoiseXTerminator, rather before.

Thank you all for your inputs. The advice to form Previews and "test, test, test" is a good one for any and all of these processes ...

Ben

I’m glad that I followed this thread, as I wasn’t aware that Graxpert had a denoising script. I’ll have to give that a try and see how it works.

That said, I’m a big fan of RC Astro products and use BTX, NXT & SXT. I also agree with several above that the BTX should be used early in the process and my workflow is as follows.

Crop edges to taste,
I use the PI Gradiant Correction tool with a couple adjustments. Can’t recall the slider names off hand (Sorry!)
From there I’ll run BTX “CORRECT ONLY”
Plate solve the image
This is kind of a personal thing but I don’t use the “estimate PSA” in BTX so I’ll convert the image to grayscale and pull the FWHM number and use that to actually run BTX. Once I have my number then I switch the image back to color.
Then I run SPCC
BTX and here about the only change I make depending on the data is how much to sharpen the stars.
Then NXT and here sometimes I’ll adjust the DeNoise slider back a bit sometimes again depending on the data and my settings are usually between 70-85% and run. I use the default “sharpening” setting.
From here I use SXT and then process the Starless and Star layers separately.

On the starless layer I will do some different things depending on the object. I like to use luminance layers to sharpen and add that back to the starless image. I’ll use different sharpening techniques depending on what results I like best which is another reason I prefer not to sharpen the image to much in either BTX and or NXT. This is just my preference and others may differ from my process.

A couple point I’ll make is I agree with the comment above “test, test, test” using the live preview box option in the tools that RC has. This make a big difference in that you can see what possible artifacts may develop with the settings you have input.

Also this is my workflow for OSC data.

I don’t know if this helps the OP’er any but this is my input to this thread and again thanks for brining to my attention the GraXpert has a Denoising script! I’ll have to give that a go.

Dale

It’s shame this thread got off the rails a while back. I was interested in the comparison in NXT between Details Slider 15 and Details Slider 0. Have to say that in general if Mr. Croman’s advice/default value is 15% that’s good enough for me. The man’s a wizard in my book. Anyway I have been getting some odd noise issues recently particularly on drizzled Galaxy images. NXT if left to it’s defaults results in harsh blotchy areas of noise in the background that are totally unacceptable. Certainly it’s something in the data but in theses cases very light stretches and protecting the background with a mask has been helpful. Then NXT after the stretch not my usual flow but it seem to work with these drizzle images.

Riedl Rudolf:
I don´t know Graxpert. I use only NxT with the following settings:
I do background fixing first, than sharpening with BxT usually in 3 steps. As someone mentioned here BxT should not be applied after Nxt. This gives bad results! I apply for the use with my both Nikon cameras (I dont use yet Astrocams) first "correct only" than I do color calibration and than the first round with setting of ar. 60% and a second round with 40%. The sharpening I set at 5%. Higher values makes the stars look too artificial at least in my configuration.
CS Rudi

I do similar processing, however I agree with you regarding star reduction artifacts with BXT when used for star reduction. Mind you, even doing sharpening only can cause some star reduction, and that is fine with me. But for anything more than slight star reduction, BXT is far inferior to Bill Blanshan's star reduction method, which is a simple pixel math function that basically works as one might do if processing stars separately without any effect on the original psf of the stars.

I will occasionally use the halo reduction of BXT slightly if desired. But never now the star reduction.

One fault of BXT star reduction is if the stars have any structure, such as even the slightest diffraction spikes (even if fully contained within the psf of a star) BXT will enhance those defects. But I hardly apply any BXT functions strongly anymore. I still find too many over-use these, even those with data from giant aperture scopes.

Many of these suites are known to create color shifts in data. I almost always use these early, as suggested, and do color calibration after. If I ever need or want to do something from this suite later, I take care and check that color calibration hasn't changed.

It’s shame this thread got off the rails a while back. I was interested in the comparison in NXT between Details Slider 15 and Details Slider 0. Have to say that in general if Mr. Croman’s advice/default value is 15% that’s good enough for me. The man’s a wizard in my book. Anyway I have been getting some odd noise issues recently particularly on drizzled Galaxy images. NXT if left to it’s defaults results in harsh blotchy areas of noise in the background that are totally unacceptable. Certainly it’s something in the data but in theses cases very light stretches and protecting the background with a mask has been helpful. Then NXT after the stretch not my usual flow but it seem to work with these drizzle images.

I have been doing a lot of wider field galaxy cluster imaging lately and noticed a fair amount of blotchy patches and regions in the background. For some my data is somewhat starved for integration time. I looked up these fields in Aladin/Simbad and found a correlation between these "artifacts" and distant galaxy clusters. It appears that the distant light from these distant galaxies can register as a smoother signal without some of the pointlike features of those galaxies. I suspect that with more data, those noise-like artifacts would reveal true structures. In fact I have recorded some of these distant clusters and partially I can see individual galaxies and surrounding and between these, I see this signal, likely light from the smaller galaxies and dim star streams, etc.

hold on a sec, let me clear this argument:

Dan Kearl:
I thought you literally meant that you don't know anyone who sets the slider to 0

Yes, I know no one who sets the detail in NXT to 0. Why do you?
If you think Russ Croman's tools are like "cheap" tools in photoshop, don't use them.
I think most Astro people appreciate his software and most use the default settings or close to them.
If you know better, fine, What is your point?
Just do your own deconvolution the manual way.

I don't do it because it seems to do more bad than good, but I'll try it out again and post a comparison for everyone to see

I didn't say I think RC tools are like the cheap tools in photoshop; I said specifically the Detail slider gave an effect similar to Photoshop sharpening

I never do deconvolution manually either; it's not like "because you don't use the Detail slider at 15, you will use manual deconvolution!" that doesn't make sense; you forgot about BXT.

I see your point if you mean that a person without the budget, who can afford only NXT, may want to use Detail to deconvolve, but I don't recommend even them to use it.

My original point was, IMO, it's not a good setting to use.

My other point, IMO, was that there was no reason for you to become aggressive.

EDIT: just to clarify, when I said "well now you know someone", I meant it in a joking way, but I had no negative intentions.

It’s shame this thread got off the rails a while back. I was interested in the comparison in NXT between Details Slider 15 and Details Slider 0.

I'll still do it

My other point, IMO, was that there was no reason for you to become aggressive.

I apologize, I took some of your phrasing the wrong way,,
That being said, I just think that people who develop software set defaults and parameters for a reason.
Of course there are sliders that go from 0-100 but there must be a good reason Croman chooses 90 and 15. I assume its not random and it suits
the algorithms he uses best.
Much like any software, Adobe Camera Raw in Photoshop chooses a WB, you can adjust it if you want but clearly they are using algorithms they feel gives the best WB.
Same with Capture One and other photo software.
Blurex has defaults, I doubt Croman made those up either and you can adjust if you want but it's not random numbers he chooses.
I tend to go with the expert and since he developed it, unless I know better I will use his.

I thought you literally meant that you don't know anyone who sets the slider to 0

I also set the nxt sharpening slider to zero. My reasoning is that the sharpening feature of nxt predates bxt so its defaults seem appropriate to a workflow wo bxt. I use bxt to sharpen / deconvolve and nxt to denoise so don’t mix functionality across the two.

would love to hear Russ’ inputs on this. He trained both cnn systems so certainly would be most authoritative here…

jim