Keep Or Discard?

[gnomus]

I have been capturing some OIII data for a Pacman image.  A couple of nights I have encountered bands of cloud rolling by.  These have not been bad enough to make me lose the guiding (indeed the guiding has remained consistently good throughout).  But there is definitely a  difference between frames taken under clear skies and skies with occasional clouds drifting by.  What I am unsure about is whether I should keep or discard those lesser frames.  I have taken two (calibrated and aligned) subframes into PI - I then did one identical histogram stretch on both, and resized for posting here.  No other processing of any sort.  I think you can see the differences in contrast.  Here is my stupid question of the day: should I aim to keep only the very best frames on the grounds that any 'impurity' will diminish the final result; or, since I have some signal and improving SNR is partly related to simple arithmetic, will inclusion of the lesser quality frames actually be better for the overall SNR of the stacked frame?  Perhaps I should point out that I have, of course, tried creating an OIII stack with (9 frames) and without (6 frames) the lesser frames.  I could not see any significant deterioration in the OIII stack with the lower quality frames within: I did think the noise was less in the 9-sub stack vs the 6-sub stack.  You might think I have answered my own question, therefore, but I am not sure if my processing of the stacks could have contributed to there being no apparent difference.     

[sloz1664]

There's only one way to find out. Do a quick and dirty full process with and without and check the results.

Steve

[gnomus]

25 minutes ago, sloz1664 said:

There's only one way to find out........

Steve

....... FIGHT!

No - I realise it was a bit of a stupid question, but I wondered if, theoretically, any signal was better than no signal.

The current version - including all (now 11) OIII frames can be found here:

[vlaiv]

I think it depends on conditions when frames were taken.

One might argue that any signal is better than no signal, but, you are not adding the signal alone. You are adding both signal and noise from lesser frame. If frame was taken in bad conditions (LP, uncooled camera - a lot of thermal noise, etc) you might end up adding more noise then signal thus hurting your final SNR. Also clouds don't only obscure target (thus lowering it's signal) but also reflect more LP than clear skies - adding more noise into the mix.

In the ideal world, knowing all the variables one would be able to calculate if there is increase or decrease of SNR when using lesser frame/s but I think that your best bet is to stack both and compare (or even do combination of lesser frames to include, some of them might increase SNR, some might decrease, depending on cloud cover density and properties, and this can be even automated, but I don't think anybody implemented such a feature in software - but there is an idea )

 

[martin_h]

I usually discard frames like that, I find they lower the overall contrast in the final stack.

[Pompey Monkey]

One of the criteria that Warren Keller mentions in his PI book for discarding subs is "lack of contrast". So that's one answer.

On the other hand, especially if you reside in Blghty, we all know how painful it is to discard that hard-won data.... 

For me, Steve's quote sums it up perfectly

[johnrt]

Have you compared the 2 stacks in PI's statistics process? That will give you an answer beyond the human eye as to what difference the poorer quality subs are making. Don't forget PI also weights the subs when stacking based on SNR, so you may find they aren't really making that much of a difference when included as PI has reduced their input.

The stacking and rejection in PI is amongst my favourite things it does, very powerful.

[wimvb]

Image 10 has significantly lower dynamic range than image 5. The more of these images you have, the lower the dynamic range in your stacked image. Only you can judge how much to tolerate.

It will depend on what you're after in your image. If you want to go deep, keep them out. If your target is bright, but you want to lower the noise, you can probably leave them in.

And then I realised that this is just a fancy way if saying: experiment. 

[swag72]

I looked into this a while ago and put the results on my website. I'd often hear the adage of 'rubbish in, rubbish out' so I did a stack of only decent stuff and then one that included the rubbish..... Didn't seem to make any discernible difference.

You can read the stuff I wrote here if you're interested

[ollypenrice]

I'm another for experimenting. You may lose contrast (you almost certainly will lose it) but this can be rcovered to some extent in processing.

There is a half way house which I once tried while imaging an insanely faint target. I made two stacks, the 'perfect stack' and 'the lot' and processed them both till they looked similar - similar background levels etc. I then pasted the Perfect Stack over The Lot and blinked it on an off. There was marginally less noise in the faint parts of The Lot so I selected and partially erased those regions which had the noise in the top.

Another thing I tried on the same project was shooting a set of luminance in Bin 1 and another in Bin 2. I resized the Bin 2, pasted the Bin 1 on top of it, selected the faint signal and partially erased it from the Bin 1. The idea was to have the bright parts at nice bin one resolution and faint parts assisted by nice low noise faint stuff in Bin 2.

Did it work? Maybe a bit, but not enough to make me go through all that bother again!!

Olly

[graemlourens]

Hi Gnomus.

I've been having this decision battle with approx 30% of my targets, as we do have lots of thin clouds here that give exactly that effect.

I ALWAYS calibrate, and register ALL images. Then i stack all of them, and i stack only the best, comparing both visually, but also with statistics in PI. When you evaluate the visual difference (that i think ist mostly not noticeable if you only have few subs that are 'bad') and also considering the noise evaluation (via script) i then decide what version to keep. It also depends what kind of target it is. Bright / Not bright and what i'm looking for in the final image (contrast or other things).

I was also tending to rubbish in rubbish out, but when you actually do the tests, it really depends on if noise or contrast is more important to you, and that really varies by target in my opinion.

Kind regards, Graem

[gnomus]

1 hour ago, swag72 said:

I looked into this a while ago and put the results on my website. I'd often hear the adage of 'rubbish in, rubbish out' so I did a stack of only decent stuff and then one that included the rubbish..... Didn't seem to make any discernible difference.

You can read the stuff I wrote here if you're interested

Thanks Sara.  That's pretty much what I was finding.  

[ollypenrice]

If you look at your pictures by measuring them in PI then believe PI. If you look at your pictures by using your eyes then believe your eyes...

I thought it looked better in version one but PI says it's better in version two...

??

Olly

[graemlourens]

9 hours ago, ollypenrice said:

If you look at your pictures by measuring them in PI then believe PI. If you look at your pictures by using your eyes then believe your eyes...

I thought it looked better in version one but PI says it's better in version two...

??

Olly

I agree totally. Its all about what you want in your result, not what statistics will tell you is best (even if mostly those should and will concur in my opinion).

[ChrisLX200]

Personally, I would discard any images that look like frame [10] because it has lost the extended faint background detail, and it won't add anything in that regard. However, what it [they] will do is reduce noise when averaged together with the better frames which would improve the high signal areas, but that is at the expense of losing the low signal. If this was to be (for e.g.,) one channel in an HST palette image then including such frames would also cause you problems with an overall colour bias requiring rather drastic steps to equalise/remove. It can be done but I find it better to avoid such issues by using consistent data. I see similar frames if I try to grab that last extra frame at the end of a session as dawn approaches, or start too early before it's got properly dark.

ChrisH

[sharkmelley]

In theory, both the good and bad frames are contributing information into the final stack.  There is no need to discard them - they just need to be optimally weighted according to how clean they are.  This weighting of exposures is something that PixInsight is able to do automatically (if it is switched on) in its integration but to be honest I've never tested how effective it is.  The only reason I discard frames is if the stars are distorted (e.g. tracking problems) or if some frames display a very obvious gradient (e.g. passing cloud, passing head torch).

Mark

[gnomus]

Having read the link referenced in Sara's post above, that is what I have decided to do too.  I don't know how to weight either.