Local Normalization and set of LRGB images

[Igor]

Hello,

I would like to start with Local Normalization and I do not know the best choices for the reference image.

With a set of L, R, G and B images, would it be better to local normalized all the images with the best Luminance image ? Or better to local normalized the set of L with the best L, the set of R with the best R , the set of G with the best G and the set of B with the best B ?

Same question with L in bin1x1 and RGB in bin2x2 ?

Thanks in advance for your replies.

Regards

Igor

[RickS]

Juan may come along and herd us gently in the right direction but I would think that it's a bad idea to use a reference from a different filter with a set of lights.  LocalNormalization is removing gradients, and if you use a red reference with a set of blue lights then it's going to interpret normal differences in the data as gradients.  I have been using a red reference for red lights, blue reference for blue lights, etc. and have been having some success...

[Igor]

Yes it sounds logical.

Thanks Rick !

[Juan Conejero]

Hi Igor,

100% agree with what Rick says. LN works for similar images. Mixing filters or filters and lights won't work.

[Igor]

Thanks Juan.

"LN works for similar images" : does it mean as well that it is better not to mix L images shot at different dates ? For example L images shot on September 7 with L images shot on September 9 ?

Regards

Igor

[drmikevt]

My experience shows that this works very well (assuming the same target and position angle). 

Although we don't know all the ins and outs of LN yet, we might argue that this is exactly what the process is for - normalizing images taken under varying sky conditions. 

In fact, attached is a little gif I made of a Blink slide show before and after applying LN.  The data is 48 Sii images taken in 2 groups, about 2 month apart.  I'll have to post the other one in a separate post for size reasons.

This one is no LN applied

[drmikevt]

And the corrected, normalized images....

[Igor]

Thanks Mike.

And regarding your 48 Sii images taken in 2 groups about 2 month apart, did you choose to local normalized your images with one reference image for your 2 groups ? Or each group (i.e. each date) with a separate reference image and separate LN process ?

[drmikevt]

Again, we are waiting for instruction, but I think the whole point of LN is to normalize your image stack to a single image.  I used the best SNR frame (also taking into account FWHM and eccentricity) from the total 48 images. 

[RickS]

I used the best SNR frame (also taking into account FWHM and eccentricity) from the total 48 images.

I use the cleanest looking sub (free of gradients) or an integration cleaned up with DBE as the reference.  SNR can be misleading.  A cloud affected sub can have great SNR, especially if the clouds are reflecting light pollution, but that doesn't make it a good reference sub.  I also suspect that FWHM and Eccentricity aren't terribly important for LocalNormalization (though they are definitely important when choosing a reference image for registration!)

Cheers,
Rick.

[drmikevt]

Yes, my apologies - I should have said that the reference needs to be inspected and DBE'd. 

But I have a question about using an integration - it would seem that, since you are using a stack of varying quality, with many of the frames being of a higher background than the frame you would choose as a reference (if you were choosing a reference that way), that the resulting integration would be of a higher-than-optimal background level due to the influence of the frames that would be most heavily normalized by LN.  This would mean (if I'm making any sense) that if your data is varied and you use an integration as the reference than the resulting, normalized integration would not have an optimal SNR due to a higher than necessary background....No?  I don't know if I'm thinking about this correctly or not, but its something I've been thinking about.  So far, for me, 'my' method (using a chosen sub) has been working well to improve the quality of the final integration over not using LN. 

Also, while we're on the topic, do you (Rick) tend to experiment with the scale parameter for difference images or do you leave it at the default of 128? 

Thanks
Mike

[RickS]

Mike,

But I have a question about using an integration - it would seem that, since you are using a stack of varying quality, with many of the frames being of a higher background than the frame you would choose as a reference (if you were choosing a reference that way), that the resulting integration would be of a higher-than-optimal background level due to the influence of the frames that would be most heavily normalized by LN.  This would mean (if I'm making any sense) that if your data is varied and you use an integration as the reference than the resulting, normalized integration would not have an optimal SNR due to a higher than necessary background....No?  I don't know if I'm thinking about this correctly or not, but its something I've been thinking about.  So far, for me, 'my' method (using a chosen sub) has been working well to improve the quality of the final integration over not using LN. 

LocalNormalization must already be making the reference and target subs statistically compatible before comparing them, so I don't think this is an issue.  An integration will have better SNR than an individual sub but that seems like an advantage rather than a problem.  The idea of using an integration came from Juan in one of his early posts about LN, so I'm pretty sure it is sound 

Also, while we're on the topic, do you (Rick) tend to experiment with the scale parameter for difference images or do you leave it at the default of 128? 

I have experimented on a few targets and for the data I have been processing a scale of 256 seems to give good results.  At smaller values I've found that the halo around stars is made more prominent which I find undesirable!

Cheers,
Rick.

[ngc1535]

This is just an opinion- but using an integrated set of images as a reference for LN seems like the most conservative/safest way.  One thing that people may not be considering is that perhaps you have a set of true Luminance images (through a "clear filter" or open) and RGB images. You *could* create an integrated RGB image, if it is cleaner, to serve as a reference for the L data set. So it is a case of using the combination of the full RGB (not a single color channel) to assist L.

Applying DBE to an image (or integrated image) to serve as a reference for LN doesn't feel right to me. It is almost like doing pre-DBE adjustments to images in a way. If you can't identify a reference image in a set of data because every frame is different- then it strikes me as an artificial adjustment (ad hoc) to use DBE to make a reference and then apply LN to do the work. You might as well use LN with the best image (or integrated set of images) from the data set and do DBE in the normal way on the result.

Perhaps these are similar routes... but one path feels odd to me.


-adam

[drmikevt]

LocalNormalization must already be making the reference and target subs statistically compatible before comparing them, so I don't think this is an issue.  An integration will have better SNR than an individual sub but that seems like an advantage rather than a problem.  The idea of using an integration came from Juan in one of his early posts about LN, so I'm pretty sure it is sound 

But this is my point - (and again, I am very open to the idea that I am not thinking about this correctly) - if LN is adjusting the frames to be more similar (statistically compatible), it must be using a reference for that adjustment, which I would assume is the reference image.  If so, I guess it comes down to:  what is the most important aspect of the reference to achieve the best resulting image?  SNR? (probably not), Noise?, Median/Background level?.  If it is background level then I still think that an integrated image *could* have an unnecessarily high background, depending on the data set.  You said earlier that SNR doesn't really matter.  If noise determines the best reference frame then, certainly, the integration is the best.  Since Juan suggested to use an integrated image, I'm sure that that is what we should do - I just want to understand the process better to know how to pick the best reference. 

Mike

[RickS]

Perhaps these are similar routes... but one path feels odd to me

Adam: I think of it in terms of finding or building the closest thing I can to a gradient free reference and I'm happy to do that either way.  I do understand what you mean though...

I'm working on an 81x600 sec luminance at present and I tried both a good quality reference sub and a DBEd integration.  I didn't find either one produced a result that was clearly better though they were subtly different.

But this is my point - (and again, I am very open to the idea that I am not thinking about this correctly) - if LN is adjusting the frames to be more similar (statistically compatible), it must be using a reference for that adjustment, which I would assume is the reference image.  If so, I guess it comes down to:  what is the most important aspect of the reference to achieve the best resulting image?  SNR? (probably not), Noise?, Median/Background level?.  If it is background level then I still think that an integrated image *could* have an unnecessarily high background, depending on the data set.  You said earlier that SNR doesn't really matter.  If noise determines the best reference frame then, certainly, the integration is the best.

Mike: my assumption is that the best reference frame is the one least affected by gradients.  LN is looking for large scale gradients so small scale issues like FWHM and noise shouldn't be significant.  I think the background level of the initial integration will depend largely on the reference image for the integration (the first one in the list of files) but I don't believe it will matter if you provide LN with a reference that has high values or low values because it will normalize the subs to the reference either way before analysing the gradients.

Since Juan suggested to use an integrated image, I'm sure that that is what we should do - I just want to understand the process better to know how to pick the best reference. 

Reading the original thread announcing the feature I came away thinking that using a reference sub was OK, using an integration was OK and that it was fair game to apply DBE to the reference as well:

https://pixinsight.com/forum/index.php?topic=11063.0

Cheers,
Rick.