Local normalization problem

P

ppeake

Active member
Spent pretty much all day working on a problem, trying to just do basic integration of a set of images.

MacOS - 10.15.3
PixInsight 1.8.8.4

Because I was getting nowhere, I tried following the recipe on the Light Vortex website, to ensure it wasn't me doing something stupid.

https://www.lightvortexastronomy.com/tutorial-pre-processing-calibrating-and-stacking-images-in-pixinsight.html

I am at the point where I have aligned the images and run local normalization on them.
The tried integrating, and got horrible results (aligned images look fine).
So backed off, re-ran the align and local normalization, then tried integrating just three images.

The integrated image is horrible:

https://vogon.net/images/Local_normalization.png

Cleared the local normalization files and switched the processing to no normalization.
Looks fine (these are 2 second exposures):

https://vogon.net/images/no_normalization.png

So, is local normalization broken? Or am i?
 
Hi,

LocalNormalization is a great tool to equalize images that are statistically different (i.e. images from different nights/equipment/atmospheric differences).
LN can work wonders and can also not contribute to the image.  Personally I have stopped using LN as I found using DBE better at removing gradients/brightness etc.

space is not black
John
 
Please see bulrichl's calibration guide for the correct way to perform data reduction in PixInsight, along with important and rigorous information on many essential topics. Instead of doing the image calibration and registration tasks manually, you can use the WeightedBatchPreprocessing script, which automates them. However, knowing the entire data reduction process and understanding how it works at each stage is absolutely necessary if you want to control your data. Automation should never come without previous knowledge of the tasks performed.

LocalNormalization is an advanced tool able to provide good improvements in difficult cases. However, this is only true when it is used correctly, and only when it is really necessary:

- Please do not use it on a regular basis, unless you have good reasons to use it. Some of them are gradients with varying orientations and relatively small differences caused by varying atmospheric conditions. If in doubt, don't use it. LocalNormalization is an advanced tool, it is not a necessary task.

- Never use it on dissimilar images. For example, don't use LN on narrowband and broadband images, or on very different images for any other reason, including significant exposure times. This does not make any sense.

- Never use it blindly with default parameters. The tool is complex and usually requires fine tuning of parameters and some trial-error work.

- Always inspect the resulting images after LN. The LN task is ill-posed by nature, so never take for granted that the implemented algorithms will succeed. The LocalNormalization tool provides many resources for quantitative and qualitative evaluation of results, which should always be used.

If you are starting, keep things simple and try to learn correct, robust procedures. You probably don't need LocalNormalization for this data set, and shouldn't use it at this point.
 
Juan Conejero said:
Instead of doing the image calibration and registration tasks manually, you can use the WeightedBatchPreprocessing script, which automates them. However, knowing the entire data reduction process and understanding how it works at each stage is absolutely necessary if you want to control your data. Automation should never come without previous knowledge of the tasks performed.
Click to expand...
My guide probably should be prefaced by this statement. I could not express it better.

Bernd
 
Juan Conejero said:
- Please do not use it on a regular basis, unless you have good reasons to use it. Some of them are gradients with varying orientations and relatively small differences caused by varying atmospheric conditions. If in doubt, don't use it. LocalNormalization is an advanced tool, it is not a necessary task.
Click to expand...

This was actually why I was trying to use it. Imaging from home, I have a town below me (not a large town, but enough to cause some light polution). Imaging for several hours, the camera traverses over the town, and I end up with gradients varying from not very much to noticeable, and at changing angles. Cleaning those up is generally a pain. I was hoping this would be a "magic bullet" - should have known that those don't really exist.

I was just surprised at how much "damage" this did though. I wouldn't have expected that.
 
i've been railing against the inclusion of LN in that light vortex tutorial in various fora for a while. while i appreciate all the tutorials and work kayron has put into his site, that one tutorial has got all kinds of people blindly inserting LN into their flow and running into unnecessary problems. juan's explanation is (of course) 100% correct; it is a tool that needs careful tuning and A/B comparisons to see if the subs were actually improved by inserting it into the flow.

anyway, you can try increasing the scale from the default of 128 to 256 to see if the blotchiness goes away. or try some factor-of-two scale between 128 and 256. try this interactively. also because these changes can be very subtle, the subs might look OK but the integration looks funny, so unfortunately you have to both test on the subs and on integrations. but the good news is if you regenerate the LN files you can just integrate over and over without much hassle.

rob
 
Juan Conejero said:
Please see bulrichl's calibration guide for the correct way to perform data reduction in PixInsight, along with important and rigorous information on many essential topics....

....LocalNormalization is an advanced tool able to provide good improvements in difficult cases. However, this is only true when it is used correctly, and only when it is really necessary:
Click to expand...

Two things: first, the link to "bulrichl's calibration guide" doesn't now seem to lead anywhere useful (at least, not that i can see) -- anyone know where it is (i'd hope that any "important and rigorous information on many essential topics" is still accessible!)?

Secondly, i arrived on this tread because i've been looking round for a comprehensible (to me...) description of what LocalNormalization actually does, both 'under the hood' AND in real-world situations -- especially in the light of the various warnings expounded above. "Extensive research" (i.e., 10 minutes googling) really hasn't led me to anything i've found useful; can anyone point me at a suitable "idiot's guide"?
Thanks
 
bulrichl said:
The link that was given above ... points to an older version of this tutorial
Click to expand...
Sorry! I was just answering the OP question of where his link had gone. I hadn't noticed that the "Tutorials and Processing Examples" version was different.
 
I wonder if someone better informed than me could explain (in a few words, not a half-hour tutorial) exactly when LN is useful. Having worked out (roughly) what it appears to do, it looks like a recipe for disaster (replacing a single consistent estimate of the statistical background across the whole frame with discontinuous local estimates across subregions of the frame).
I must point out that my interest is academic - I have never actually used it (I don't like using tools I don't understand)!
 
it's supposed to do what it says... normalization in a local fashion. this would be useful in the presense of gradients, for instance. it does require a lot of care to get it right but it has helped me from time to time as i do image from bortle 8-9 skies. it's kind of a miracle when you end up with a mostly gradient-free image from lousy subs.

rob
 
My urban backyard site always produces LP gradients. In addition, the UK weather means I have to take my chances when I can, which seems to be strongly correlated with the full moon ... so yet more gradients. I have recently been having increasing success with DBE, once I realised that:
  • I really pays off to use only manually placed samples - lots, very carefully positioned, and
  • with strong gradients it is often worth curing the worst with one application, and tidying up the residue with a second.
 
fredvanner said:
My urban backyard site always produces LP gradients. In addition, the UK weather means I have to take my chances when I can, which seems to be strongly correlated with the full moon ... so yet more gradients. I have recently been having increasing success with DBE, once I realised that:
  • I really pays off to use only manually placed samples - lots, very carefully positioned, and
  • with strong gradients it is often worth curing the worst with one application, and tidying up the residue with a second.
Click to expand...
I have similar problems with both UK weather and very strong local light pollution gradients!

I am currently writing a script 'NormalizeScaleGradient'. The reference image should be the one with the least gradient - probably when the object was highest in the sky. All target images will be scaled to this reference image, and their relative gradient will be removed. The strategy used is very similar to PhotometricMosaic.

1618260714239.png
 
i've ended up with gradients so crazy in the integrated images that DBE is pretty difficult. it's never impossible - just hard - and so on a few occasions it made more sense to use LN such that the integrated gradient was a little less complex.

i look forward to @jmurphy's script! any new tool to help with this is welcome.
 
You must log in or register to reply here.
Back
Top