[Sticky] What exactly does LNC or Local Normalization Correction and when do I change the default values to something else?

Hi Frank,

LNC stands for Local Normalization Correction.

It will correct the illumination in your frames in such a manner that the illumination differences of the separate frames in your stack will be matched locally instead of globally.

Normally, normalization is only corrected globally per frame. With LNC the gradients in  your data will be matched per frame while maintaining a solution for all frames at once that is stable and as flat as possible.

If you are combining data shot at 1 single night then LNC will still help, but the improvement is very little and visually hard to see, because locally the frames already match pretty well.

So if you are combining data with deviating gradients (multiple imaging sessions, data from different setups and/or photographers) then LNC should give a clear improvement.

An example: data of the Cocoon Nebula shot with a Nikon D5100 and a Robtics 102 ED apo. 2 frames shot at different nights with different moon fase and light pollution gradients and slightly different Field of Views. The effect of LNC is best seen on the borders where the stack artefacts are still visible.

Top row: neutralize background disabled in 5)NORMALIZE

Bottom row: neutralize background enabled in 5)NORMALIZE

Left: no LNC

Middle: LNC 1st degree 1 iteration

Right: LNC 4th degree 3 iterations

When to use which settings:

For regular integrations (I mean not a mosaic), start with 1st degree and 1 iteration. Then improve from there if needed. First increase iterations to 3. If that doesn't change the outcome, increase the degree. For regular integrations (no mosaics) normally 2nd or 4th degree with 3 iterations gives a very nice improvement.

For mosaics, it really depends on the data which degree and the amount of iterations that are needed. The user should experiment, but I recommend, just like for regular integrations, start with 1st degree and 1 iteration and then improve from there.

Mabula

p.s. I have made this a sticky 😉

Hi Mabula, 

would You please provide some more information about the degree parameter? Some questions:

- What is the difference between 1st vs. 2nd vs. 4th degree etc.?

- How are the factors  for performance / calculation time required by the various degrees? I tried 4th degree the other day and after more than 1 hour calculation for the LNC parameters I needed to abort (2nd degree iteration is usually half an hour or so on my machine. Is this to be expected?

- Is there any difference in RAM needed or HDD space required between the degrees?

Thanks in advance,

Jochen

Posted by: @jochen-scharmann

Hi Mabula, 

would You please provide some more information about the degree parameter? Some questions:

- What is the difference between 1st vs. 2nd vs. 4th degree etc.?

1st degree means that only linear adjustments are done on all images, 2nd means kwadratic adjustments and so on. The correction models are bivariate polynomials :

https://en.wikipedia.org/wiki/Polynomial#bivariate

- How are the factors  for performance / calculation time required by the various degrees? I tried 4th degree the other day and after more than 1 hour calculation for the LNC parameters I needed to abort (2nd degree iteration is usually half an hour or so on my machine. Is this to be expected?

Yes, the higher degree the more parameters need to be solved per frame and the complexity rises fast per degree. So yes 4th degree takes more than double the time of second degree.

- Is there any difference in RAM needed or HDD space required between the degrees?

Yes, you need more RAM with a higher degree and with more frames. The Matrices that need to be created and calculated grow with higher degree and more frames. In the current LNC implementation running 1000 frames with LNC2 requires a serious amount of RAM...

Going forward though, soon I will be able to make the LNC calculation faster I think and I will revise/update the LNC algorithms as well. Currently, LNC only runs on 1 cpu thread and I will be able to make it run using multiple threads. So that will speed it up considerably. And by revising the algorithms I should be able to make it smarter and run with less memory.

Mabula

Thanks Mabula, 

Glad to hear this will be getting faster by multithreading - integration and LNC especially is the most time consuming activity currently.

is there a way to skip remaining LNC iterations?

Given the time needed per iteration it would helpful to have a "skip further iterations" button or an automated skip once the latest correction is lower than a certain threshold. I noticed LNC stops on its own already once the remaining error reaches 0 in all channels the second time with the message "no further improvement. But once 0,  it could just stop because anyway no further improvement is to be expected.

 What would be a reasonable value for a threshold?

cheers,

Jochen