Not seeing any improvment in SNR with Linear Fit Clipping over WS Clipping

[astrovienna]

Shouldn't the Linear Fit Clipping result be less noisy than the Winsorized Sigma Clipping result?    I'm running 92 x 20 mins Ha frames through Image Integration.  I've cranked the low/high rejection settings all the way to 5/10 (I only need 4/4 with WS clipping) and now the LFit result looks(*) very close to the WS result in noise, although the NoiseEvaluation script shows the WS clipping version still has very slightly lower noise (45.55 vs 45.75).  All of the other settings are identical.

This is a very faint target (planetary nebula HDW2).  The brightest part of the nebula has a K reading of 0.0061 in the integration.

(*) As far as eyeballing the results go, I'm using a saved STF icon to be sure I'm applying the same STF settings to both images.

Kevin

[mschuster]

I think it is safe to say no, well maybe no.

Linear Fit ends up rejecting fewer pixels, fewer outliers will be considered outliers. With fewer pixels rejected there might be more variance (more scale, more dispersion, etc). So anything that measures these aspects may see an increase I think.

Thanks,
Mike

[vicent_peris]

Hi Kevin,

The main goal of a rejection algorithm is not to improve noise level but to reject outliers. Therefore, usually the main reason to use linear fit clipping instead of winsorized is to be able to apply a better rejection of the artifacts without rejecting valuable data. That being said, linear fit clipping usually will improve the noise level a tiny bit.

The key is always to look at the rejection maps.

And this is always a strong reason why we shouln't care about the maximum number of subframes but about the minimum. Below 15 subframes you won't get any benefit of using linear fit clipping and, believe me, in most cases you will want to use this rejection algorithm.

Merry Christmas,
Vicent.

[astrovienna]

Vincent, thanks as always for the detailed information.

Mike, thanks as well.  And thanks especially for Mure Denoise. I finally put some dedicated effort into learning it for this image, and it's really impressive.

Kevin

[astrovienna]

I'm reviving this older thread because I've done some more testing, and I don't understand the impact of low range clipping on PixInsight's SNR calculations.  I think this explains the unexpected results I saw above, and again today.

I've been comparing the algorithms to integrate a set of 27 calibrated OIII frames of Abell 72, a faint PN.  (I previously corrected hot pixels with CosmeticCorrection, using a master dark created from 90 darks with the same temperature and exposure time of the lights.)  I created three stacks of the calibrated lights, one with no pixel rejection, one with WSC rejection, and one with LFC rejection.  I found I got the most satisfactory pixel rejection with low/high settings of 4/3 on WSC and 6/6 on LFC.

The process console report of the noise reduction depended greatly on whether Clip Low Range was checked or not.  Here are the results:

Clip Low Range Unchecked
No rejection (reference/median noise reduction): 1.4542/1.6200
LFC66:  1.3167/1.4668
WSC43:  1.3091/1.4584

Clip Low Range Checked
No rejection:  1.3205/1.4711
LFC66:  1.3174/1.4676
WSC43:  1.4495/1.6147

With Clip Low Range inactive, the LFC result has more noise reduction.  With Clip Low Range active, the WSC version has more.  Can anyone help me understand the results?  My sense is that I should have Clip Low Range active to eliminate dead pixels, but I don't see why that would change the relative SNR between the WSC and LFC versions. 

Kevin