NosieXterminator and Autostretch

I think it is stretching it harder as the noise is lower, same thing happens when you do linear noise reduction in PI

Roger,

I have not had any problems with "star rings".  24-bit lookup table is activated by clicking the "24" icon on the upper right of the ribbon near the STF reset buttons.

Jim

As many know there are many ways to do noise reduction in PixInsight.  NoiseXterminator is another wonderful tool and probably the easiest to use.  Since it uses a neural network that was trained on many types of images, its only as good as the images that it was trained on.  If you find an image that it does not work well on consider sending Russ Croman the image in xisf format.  He may add it to the next training run.  I talked to him last week at AIC and he was asking for those types of images. 

That said, the auto screen stretch stretches the image based on the statistics.  After you have run any type  of noise reduction the noise statistics change which is why "auto" gives a weird and sometimes aggressively nasty result.  I've find that its easiest to just use the screen transfer function and screen stretch your image to a point which you like it manually.  Auto gives a start then adjust from there.  You can also bring up the histogram transfer and move your screen stretch to the histogram tool and see what it looks like.  If you've clipped too many pixels the histogram tool will show you.  You can make an adjustment there and transfer that back to the screen transfer function tool by dragging the blue arrow down to the bottom of the tool.  Once you get the image looking like you want run noise exterminator.  As others have mentioned you don't want a totally smooth background.  Leave just a bit of graininess to it as it make the overall image look a bit sharper and better.  If you find noisexterminator knocks out some detail in an area of interest you can also always create a mask using something like the "Game" script and put the mask over the area to limit how much noise reduction takes place.  A simple luminance mask works well too, just remember to invert it.

Jon

Hi,

My trial version of NoiseXterminator is about to expire and I have a question before purchasing a permanent license. I'm really happy with how well it works, but I think I am using it incorrectly. I have been using it on OSC images early in my workflow (based on Croman's website recommendation) right after dynamic background extraction. However, once I perform the operation--using 24-bit lookup tables--and reset the screen transfer function and then reapply the STF, I get all sorts of color blotches all over the image. Am I not supposed to use autostretches after applying NoiseXterminator? Thanks for any help you can provide on what I hope is an easy question to answer.

Jim

I purchased the license.  Here is the STF trick I use.  Before I do the noise extermination on the linear image I get the STF reading.  After the extermination, I apply the STF taken BEFORE extermination to the exterminated image.  Seems to stretch correctly.... YMMV.

This is tricky as I have not developed the optimal workflow but I am in hook line and sinker to this ecosystem, even after buying $$$Topaz. 

I am now just doing a NB RGB image and in the RGB I am first stretching the image with ArcSin stretch, then I am getting the stars with StarX, then I apply NoiseX to the stars... again YMMV... I also try using Dust and Scraches at 1 or 2 on the stars to get rid of the fuzz... but it is minimal. 



Interesting, that when I denoise I get a warning that it thinks the ArcSin stretch image is linear... does nothing on the small scale crud..

Photoshop:




Hey, am sticking with PS on this noise reduction problem...

I'll have Sh-2 112 out soon... tune in for the NB_RGB image....

Yes, I think the noise stats change after NoiseX, but the stretch(parameters)  should be the same.  The other writer was I think referring to so big changes if you try to run the Auto STF again after noise reduction.  I have had problems with that too.  I have recently run across Generalized Hyperbolic Stretch, which seems pretty powerful for the initial stretch but has a substantial learning curve.  I have never had much luck manually adjusting STF either.  GHS I sorta understand the theory at least. 

I see you point on running NoiseX on stars.  That is I am afraid where PS Dust and Scratches is coming in for me. The stars generated by StarX seem better than any other alternative despite the small speckles, which Rob Pfile thinks are a masking artifact.  I was unfamiliar with ACDNR but it seems to provide a similar clearing effect with small structure size. 

@jerryyyyy , @Jon Talbot ,

Thanks for the tips. I got the STF readings from the pre-noise reduction image and then applied those readings to the image after NoiseXterminator and the stretch worked correctly.

@Roger Menard I use the same process and have not had any problems so far.

Thanks all for the quick responses,

Jim

I am still having problems with NXT. It does a great job of removing noise, but creates artifacts around stars.
Here is an image of M85. It has been through the processes:
Flat cal, FLat integration
Image Cal, Image Debayer, SubframeSelector, Local Normalization, Star Alignment, Image Integration.
While still in linear mode I do DBE, then use NXT. I then went directly to ArcSinH stretch (I normally would od a PCC but I am just checking out noise so I skipped this step)

I then cropped the image to clearly show  the stars..  I previously processed the same image with the same steps using EzDenoise and did not get the star artifacts.

Is there an adjustment to NXT that I should be trying. I currently use both defaults. Clearly the NXT does a great job with noise... but does impact the stars.

I had some bad issues with SXT and a particular image and had an extensive email discussion with Russ Croman.  NXT should not be applied early in process, instead late if not the last thing done.

That is because it has a significant effect on the PSF of the stars/image which in turn effects the AI in BXT and SXT (and not in good ways).  The effects are exacerbated in auto stretch, particularly in dense star fields.

Russ is very reluctant to make any recommendations to anyone on workflow/workflow order since that is a topic that is a religion to some imagers.

However, I went from the early often camp on NXT to late/last and have had consistently good results, much better and consistently that I was having previously.  SPCC is another process that is intensely affected by anything that changes the PSF of the image, to the point that NXT can cause SPCC to fail in its plate solving attempt.  Where SXT is concerned, the effects on PSF from NXT can degrade SXT's ability to identify star vs other objects like galaxies.

BXT does work with SXT, so the order between the two isn't super critical, but I've been applying BXT prior to SXT with good results.

Another process that can negatively affect SPCC and SXT is drizzle in WBPP.  According to Russ, the 1x setting is manageable, but 2x or higher can degrade the performance of both.  I conducted several test WBPP/work flow process/order tests in the course of our discussion and gave Russ the results.  

My updated workflow after my exchange of information/images/discussion with Russ:
DBE - Color Correction/SPCC - BXT - SXT - the work flow splits here for the Stars images vs starless that result from application of SXT:
Stars - Histogram stretch to get the desired intensity, then Color Sat - CorrectMagentaStars script - optional SCNR.

Starless - STF (my tool of choice has become Bill Blanshards PixelMath stretching scripts) - the whatever the image requires in the way of masks, CRV tool - etc.  The last or next to last step is application of NXT.  

Then test runs of recombining the Stars + Starless frames, etc to get the final image.

Another reason for waiting until late/last in the processing to apply NXT is that almost every post processing step after STF has some effect on visual noise/graininess in the image - not in a good way.  Applying NXT at the end smooths that graininess/noise away.  

For context, my primary goal was always to get a clean, high quality Stars and Starless image separation.  I always process my Star field separately from that of the main target, even with images like M31 or M101.  Stretching and tweaking the stars image is quick and simple while the image of the nebula, etc can require a lot of steps that would adversly affect the quality of the star field part of the image.

Early application of NXT (excluding the other issues it creates) will mandate multiple applications of NXT to deal with the process generated noise in the later stages.  I've demonstrated to my satisfaction, that multiple applications can introduce artifacts around stars and nebula detail just as it does with other noise mitigation processes, including Topaz noise reduction outside PI.

 YMMV.

CS
Clayton

Jan,
I need to qualify my comment above very slightly.

Since June of last year my imaging has been solely monochrome (LRGB or narrowband).  Since going to monochrome image noise has gone from a significant issue (in my OSC imaging) to an almost negligible issue with the primary source being post processing induced noise.  Which, aside from the superior image detail, is another major reason why I won't go back to OSC imaging.

Which points out a step I didn't put in.  For monochrome, there is a channel combination between DBE and SPCC/ColorCal.

IMO - YMMV

I have BXT (BlurXterminator) in my flow.  BXT does some pretty significant (usually good) things to images and it doesn't conflict/create insurmountable issues with the performance of SXT that I can tell.  According the the interviews and discussion with Russ Croman, BXT's AI "looks" at the entirity of the image so works best when channel combination has been completed.  ColorCalibration doesn't bother it either.

The auto function works well, but it works best when you determine the average PSF from the image in an area devoid of nebula/other objects.  Russ demonstrates this using the PSF function in PI while manually selecting stars.  However, VisibleDark demonstrated in his video a (for me) better/faster method.  
1) Create a preview pane of a small subset of the image, clear of nebula etc that has a fairly good star population but not any abnormally bright ones
2) Make sure that the preview is selected/has the focus in your image pane
3) Script/Render/PSFImage (it auto selects the image with the focus), execute with the defaults; its display will give you two FWHM values (x and y).  Split the difference and enter that value in the manual PSF value in BlurXterminator.
4) Select "Correct First" in the BXT options and execute.

You can use the BXT preview or create a preview image to apply it to while tweaking the other settings.  I tried that a few times and there may be some particular objects/images that need that increased tweaking, but I haven't found that it produces a result better than the defaults, so that is where I've left mine set other than the PSF input.

One of the reasons this works better in tuning for your image is that the PSFImage script is restricted to a specific clear/representative area without nebula features or super bright stars.  On the Auto setting, BXT's AI is looking at the entire image and trying to identify/exclude the things that would skew the result, but it will never be as 'focused' as result as human intervention can provide.  Russ Croman discusses the details of BXT and ways to get the best out of it in a video on the AstroImagingChannel YouTube channel.

He does point out that NXT changes the PSF of an image enough that BXT's AI maybe reduced to a simple sharpening routine because it cannot sufficiently differentiate stars for other objects.  Which is the same issue that SXT has.

I have found that with BXT, I really don't need drizzle to reduce star pixelization, nor do very minor tracking artifacts (stars very slightly oblong) persist into the final image.  I believe that Russ mentions this in one of the interviews/discussions on YouTube he participated in.

If any of you are still following these issues made some comments in this recent image:


Sh2 90