#1  
Old 25-04-2016, 10:31 AM
Shiraz's Avatar
Shiraz (Ray)
Registered User

Shiraz is offline
 
Join Date: Apr 2010
Location: ardrossan south australia
Posts: 4,918
Comparison of colour exposure strategies:

There are many ways to expose colour images using filters and a mono camera. Some methods seem to be significantly more efficient than others (time wise). The attached spreadsheet shows the results of a very simplistic analysis based on scaling to get to the same SNR as the first strategy presented. Time in hours has been presented for my system for a bright target, but the main point is the comparison of the different approaches, not the absolute time.

The four possibilities summarised in the attached spreadsheet are:
1. Lum + RGB full res where the RGB subs are not long enough to be sky limited – ie they cannot be added to the luminance stack, but may be adjusted for the desired dynamic range.
2. As above but with RGB subs that are also sky limited and that are added to the luminance stack. This approach does not allow the RGB subs to be deliberately shorter to increase the colour dynamic range.
3. Lum + 2x2binned RGB where the RGB data only requires ¼ as long, but has lower resolution, so cannot be added to the luminance – question is, does this lower res introduce any complications?
4. RGB with subs long enough to be sky limited, but with no luminance. This cannot be equivalent to 1. above, so have chosen the time required to get to the same luminance SNR, but with higher quality colour results. It can take a long time to get results, but maybe they are worth the wait?
5. added on Rick's suggestion in later post - RGB 1x1 subs long enough to add to luminance, then software bin RGB to generate the colour data.

Based on this very simple comparison, there may be significant gains to be made in time required for an image, simply by choosing an appropriate strategy. I have been using strategy 3. above with reasonable results and it is definitely quicker to get to a given result than my previous use of strategy 1. Would be grateful for any feedback/opinion/corrections – a simplistic analysis like this is only just a starting point.
Regards Ray
Attached Thumbnails
Click for full-size image (comparo.jpg)
69.7 KB44 views

Last edited by Shiraz; 26-04-2016 at 11:42 AM.
Reply With Quote
  #2  
Old 25-04-2016, 10:48 AM
codemonkey's Avatar
codemonkey (Lee)
Lee "Wormsy" Borsboom

codemonkey is offline
 
Join Date: Jul 2013
Location: Kilcoy, QLD
Posts: 2,058
Quick thought Ray; have you considered strategy #3 in combination with drizzling?
Reply With Quote
  #3  
Old 25-04-2016, 10:55 AM
Shiraz's Avatar
Shiraz (Ray)
Registered User

Shiraz is offline
 
Join Date: Apr 2010
Location: ardrossan south australia
Posts: 4,918
Quote:
Originally Posted by codemonkey View Post
Quick thought Ray; have you considered strategy #3 in combination with drizzling?
thanks Lee. No I hadn't - not sure how to handle SNR with drizzle, but it is definitely an option. regards
Reply With Quote
  #4  
Old 25-04-2016, 11:09 AM
Merlin66's Avatar
Merlin66 (Ken)
Registered User

Merlin66 is offline
 
Join Date: Oct 2005
Location: Junortoun Vic
Posts: 8,904
Ray,
Obviously we adopt different measures to evaluate SNR in our spectra....
What process do you use for AP images?
(I haven't used it for a while, but I believe Imagej can analyse the image for background SNR).
The sensitivity of the OSC pixels is a mixture of the Bayer filter bandwidth and transmission, the underlying QE of the chip and the eventual De-Bayering algorithm used ( i.e. Bilinear v's VNG (variable number of gradients))
Reply With Quote
  #5  
Old 25-04-2016, 11:21 AM
Camelopardalis's Avatar
Camelopardalis (Dunk)
Drifting from the pole

Camelopardalis is offline
 
Join Date: Feb 2013
Location: Brisbane
Posts: 5,425
Perfect timing Ray as I've just got myself my first mono chip

My first experiment has been with luminance subs at 1x1 and then RGB subs at 2x2 binning of the same exposure time as luminance. I think that at the 1x1 resolution of my experiment, 2.4"/pixel, the colour data may not be fine enough and many stars are clipped.

I figure I will try 1x1 for RGB next...but given 1/3 the photon flux, the colour exposures would need to be longer than the luminance, unless I misunderstand.
Reply With Quote
  #6  
Old 25-04-2016, 11:52 AM
codemonkey's Avatar
codemonkey (Lee)
Lee "Wormsy" Borsboom

codemonkey is offline
 
Join Date: Jul 2013
Location: Kilcoy, QLD
Posts: 2,058
Quote:
Originally Posted by Shiraz View Post
thanks Lee. No I hadn't - not sure how to handle SNR with drizzle, but it is definitely an option. regards
I was mainly thinking in terms of 2x2 binned RGB "adding complications" as you put it; specifically I was thinking about the annoying artefacts I get around bright stars which is worsened when binning RGB.

Another interesting option might be the possibility of being able to combine the drizzled RGB with the luminance. I have NFI if it'd work out better, but food for thought. Think I'll give it a shot soon.
Reply With Quote
  #7  
Old 25-04-2016, 01:21 PM
Camelopardalis's Avatar
Camelopardalis (Dunk)
Drifting from the pole

Camelopardalis is offline
 
Join Date: Feb 2013
Location: Brisbane
Posts: 5,425
Interesting Lee, I'll have to give it a try! I seem to get a coloured ring around stars. It's possible of course I didn't check the focus after changing filters
Reply With Quote
  #8  
Old 25-04-2016, 01:27 PM
RickS's Avatar
RickS (Rick)
PI cult recruiter

RickS is offline
 
Join Date: Apr 2010
Location: Brisbane
Posts: 10,584
An interesting comparison, Ray.

A few quick thoughts:

With option 3 are you assuming an improvement in read noise from hardware binning? If you're sky limited then software binning should work nearly as well and you'd be able to add the RGB data in to luminance.

Have you thought about the impact of noise reduction? It seems to me that binning is really a dumb form of NR. Dumb in the sense that it blurs all the data, not just the low SNR areas.

I have used option 4 with bright, colourful targets, like M31, and it works brilliantly in that situation.

Wrt Drizzling, it does complicated things with noise (see http://www.stsci.edu/hst/HST_overvie...34.html#385457 if you're interested). In practice, I think it's effectively a noise reduction technique as the correlated noise is less obvious visually.

Cheers,
Rick.
Reply With Quote
  #9  
Old 25-04-2016, 04:06 PM
Shiraz's Avatar
Shiraz (Ray)
Registered User

Shiraz is offline
 
Join Date: Apr 2010
Location: ardrossan south australia
Posts: 4,918
Quote:
Originally Posted by Merlin66 View Post
Ray,
Obviously we adopt different measures to evaluate SNR in our spectra....
What process do you use for AP images?
(I haven't used it for a while, but I believe Imagej can analyse the image for background SNR).
The sensitivity of the OSC pixels is a mixture of the Bayer filter bandwidth and transmission, the underlying QE of the chip and the eventual De-Bayering algorithm used ( i.e. Bilinear v's VNG (variable number of gradients))
Hi Ken. I would assume that the basic idea is the same, but the implementation would be entirely different, since an image is not a fixed thing - the signal varies with which bit of the image you choose. However, shot noise from the background sky glow should entirely dominate the noise and that is easy. Standard techniques for measuring image SNR involve the correlation between two images of the same scene - the correlated bits are the signal and the uncorrelated bits relate to the noise. However, it is rare for anyone to bother measuring image SNR, since the image appearance is ultimately what matters and the appearance of an image is definitely related to SNR. Re measurement, I have used Nebulosity to select specific 29x29 regions in images and measured local mean and variance for direct SNR comparison (on synthetic images, some of the central region of Helix and an outlying region in 2997). However, for comparative analyses like this one, knowing how the SNR scales with the various parameters is all that matters - absolute SNR would add nothing.

Quote:
Originally Posted by Camelopardalis View Post
Perfect timing Ray as I've just got myself my first mono chip

My first experiment has been with luminance subs at 1x1 and then RGB subs at 2x2 binning of the same exposure time as luminance. I think that at the 1x1 resolution of my experiment, 2.4"/pixel, the colour data may not be fine enough and many stars are clipped.

I figure I will try 1x1 for RGB next...but given 1/3 the photon flux, the colour exposures would need to be longer than the luminance, unless I misunderstand.
sounds like a plan to me. If the stars are clipped, maybe reduce the exposures - as a first go, possibly try 1x1 RGB at the same exposure as lum and work up from that?
FWIW, I have been using 2x2 binning for RGB with 180sec for RG and 300s for B - to go with lum 1x1 at 300s.

Quote:
Originally Posted by codemonkey View Post
I was mainly thinking in terms of 2x2 binned RGB "adding complications" as you put it; specifically I was thinking about the annoying artefacts I get around bright stars which is worsened when binning RGB.

Another interesting option might be the possibility of being able to combine the drizzled RGB with the luminance. I have NFI if it'd work out better, but food for thought. Think I'll give it a shot soon.
let us know how you get on. If using PI, You may be able to fix the artefacts around 2x2 stars by using something simple (eg bilinear) as the registration interpolation function or changing the clamping threshold (if sticking with the default).

Quote:
Originally Posted by RickS View Post
An interesting comparison, Ray.

A few quick thoughts:

With option 3 are you assuming an improvement in read noise from hardware binning? If you're sky limited then software binning should work nearly as well and you'd be able to add the RGB data in to luminance.

Have you thought about the impact of noise reduction? It seems to me that binning is really a dumb form of NR. Dumb in the sense that it blurs all the data, not just the low SNR areas.

I have used option 4 with bright, colourful targets, like M31, and it works brilliantly in that situation.

Wrt Drizzling, it does complicated things with noise (see http://www.stsci.edu/hst/HST_overvie...34.html#385457 if you're interested). In practice, I think it's effectively a noise reduction technique as the correlated noise is less obvious visually.

Cheers,
Rick.
Thanks Rick. That strategy is even smarter - get the best of all worlds - probably need to add in a few short RGB subs to boost the dynamic range - will add it the table.

I chose 2x2 binning as the standard spatial filter simply because it is clear what it does. I know that the hardware implementation leaves a bit to be desired on some chips, but that can be worked around. Other filters would be better, but I am not sure how to quantify what they do to the SNR.

thanks for the drizzle link - again, it doesn't fit easily into a simple analysis like this, but would definitely be worth some follow-up experimentation.

thanks for the input folks. Looks like a bit of experimentation is in order. regards ray

Last edited by Shiraz; 25-04-2016 at 05:10 PM.
Reply With Quote
  #10  
Old 25-04-2016, 04:32 PM
codemonkey's Avatar
codemonkey (Lee)
Lee "Wormsy" Borsboom

codemonkey is offline
 
Join Date: Jul 2013
Location: Kilcoy, QLD
Posts: 2,058
Quote:
Originally Posted by RickS View Post
With option 3 are you assuming an improvement in read noise from hardware binning? If you're sky limited then software binning should work nearly as well and you'd be able to add the RGB data in to luminance..
Very good point, Rick. I actually tried this once and the numbers indicated the SNR was better in the software binned version. I wasn't sure how to measure the "sharpness" of the images though, so it could just be that the software binned version lost some edge contrast thus reducing the variance?

Quote:
Originally Posted by Shiraz View Post
let us know how you get on. If using PI, You may be able to fix the artefacts around 2x2 stars by using something simple (eg bilinear) as the registration interpolation function or changing the clamping threshold (if sticking with the default).
There's a thought. I've messed around a bit with different integration algorithms/parameters, but I never thought to look to registration as a solution for that. Thanks Ray, I'll have to dig into that.
Reply With Quote
  #11  
Old 25-04-2016, 05:23 PM
RickS's Avatar
RickS (Rick)
PI cult recruiter

RickS is offline
 
Join Date: Apr 2010
Location: Brisbane
Posts: 10,584
Quote:
Originally Posted by codemonkey View Post
Very good point, Rick. I actually tried this once and the numbers indicated the SNR was better in the software binned version. I wasn't sure how to measure the "sharpness" of the images though, so it could just be that the software binned version lost some edge contrast thus reducing the variance?
I use FWHM as a proxy for sharpness, Lee. It's probably not perfect but it's better than an eyeball!
Reply With Quote
  #12  
Old 25-04-2016, 06:06 PM
Shiraz's Avatar
Shiraz (Ray)
Registered User

Shiraz is offline
 
Join Date: Apr 2010
Location: ardrossan south australia
Posts: 4,918
Rick's strategy added to original post
Attached Thumbnails
Click for full-size image (comparo.jpg)
69.7 KB26 views

Last edited by Shiraz; 26-04-2016 at 11:42 AM.
Reply With Quote
  #13  
Old 26-04-2016, 09:03 AM
JohnH's Avatar
JohnH
Member # 159

JohnH is offline
 
Join Date: Feb 2005
Location: NSW
Posts: 1,226
LRGB exposure times?

Your table for entry for "L with RGB binned 2x2" has exposure times of 6 (lum) and 2 (RGB) total 8 hrs. But binning 2x2 gives you 4x sensitivity so the exposure should be 1/4 of lum not 1/3, thus it should be 6 hrs Lum and then 1.5hrs each for R, G and B - total 10.5 hrs for a shot with the same s/n as one RGB at 6,6,6 - 18 hrs total or at least that is how I understand it.
Reply With Quote
  #14  
Old 26-04-2016, 11:38 AM
Shiraz's Avatar
Shiraz (Ray)
Registered User

Shiraz is offline
 
Join Date: Apr 2010
Location: ardrossan south australia
Posts: 4,918
thanks John. You are right that the binning will give 1.5 hours rather than 2 hours - the table has been changed to reflect that - thanks for the correction (I hope that it is now right). However, for the 1x1 binning in strategy 1, the total RGB time given is 6 hours (2hrs each), so I think that the total RGB time with 2x2 binning is 1.5 hours (0.5 hrs each rather than 1.5 hours each) and the total time is 7.5 hrs rather than 10.5hrs.

The luminance SNR of the RGB strategy will be determined by the total number of photons detected. Since each RGB will have ~1/3 the photon flux of the full broadband luminance (due to the filters), the total number of photons detected in any given time will be ~1/3 of those detected broadband - ie you need 3x the time to get to the same luminance SNR from the RGB strategy as you would with full luminance broadband. The colour data from the RGB strategy will of course have much better SNR than the colour from strategy 1. The RGB strategy cannot be made directly comparable with the others in this type of comparison, so I specified that the luminance SNR should be equivalent - with the note that the colour data will be better than that from other strategies. I have no idea how much difference the extra colour SNR makes, but many people use straight RGB very effectively. I guess the other thing that might be important is that RGB imaging will give better resolution than lum+RGB if the scope has any CA - may be important for some refractor users.

Once again, thanks very much. ray
Attached Thumbnails
Click for full-size image (comparo.jpg)
69.7 KB34 views

Last edited by Shiraz; 26-04-2016 at 02:22 PM.
Reply With Quote
  #15  
Old 27-04-2016, 06:40 AM
gregbradley's Avatar
gregbradley
Registered User

gregbradley is offline
 
Join Date: Feb 2006
Location: Sydney
Posts: 17,877
1x1 for luminance and 2x2 for colour has been my most used approach. Its the most time effective and little downside.

I do use 1x1 for everything for galaxies when I want detail. But even there if time is a constraint I go for 2x2 on rgb.

If the seeing is good then 1x1 everything makes sense. Seeing is a great evener.

Small pixels like in 694 sensor make 2x2 practical at times especially for narrowband unless it great seeing again.

Very often the practical limit on our imaging is available time given how infrequently clear skies occur on no moon nights around major cities.
So anything that gets higher SNR in less time like more sensitive cameras, faster scopes makes a lot of sense.

Greg.
Reply With Quote
  #16  
Old 27-04-2016, 08:03 AM
multiweb's Avatar
multiweb (Marc)
ze frogginator

multiweb is offline
 
Join Date: Oct 2007
Location: Sydney
Posts: 22,060
Good comparison Ray. Bookmarked.
Reply With Quote
  #17  
Old 27-04-2016, 03:06 PM
troypiggo's Avatar
troypiggo (Troy)
Bust Duster

troypiggo is offline
 
Join Date: Nov 2008
Location: Brisbane, Australia
Posts: 4,846
Great thread. I have mostly imaged luminance and narrowband at 1x1, and RGB at 2x2 (method 3 above). Recently experimented with RGB only at 1x1 (method 4) and also added some lum 1x1 (method 1) but the additional time (which is at a premium between clouds) was a killer for me.

I'm confused by "Rick's method" of software binning. What's the actual process here? Image capture at 1x1 for L and RGB I assume. But what exactly is "software binning" and how would that benefit you in terms of image capture time? Must be missing something here.
Reply With Quote
  #18  
Old 27-04-2016, 04:20 PM
RickS's Avatar
RickS (Rick)
PI cult recruiter

RickS is offline
 
Join Date: Apr 2010
Location: Brisbane
Posts: 10,584
Quote:
Originally Posted by troypiggo View Post
I'm confused by "Rick's method" of software binning. What's the actual process here? Image capture at 1x1 for L and RGB I assume. But what exactly is "software binning" and how would that benefit you in terms of image capture time? Must be missing something here.
In PixInsight you'd do software binning with IntegerResample (downsample by 2 in Average mode). As with hardware x2 binning you end up with 1/4 of the pixels but twice the per-pixel SNR.

The only difference between hardware and software binning is that you may get a reduction in read noise with hardware binning. If you're sky limited or have a sensor that doesn't bin well (looking at you KAF-*) then this is pretty much irrelevant.

I wouldn't actually do what I suggested. The software binning is really just acting as a noise reduction technique and there are better ways to do that. Since noise tends to be clustered at small spatial scales and low brightness levels, something like MultiscaleLinearTransform with a linear mask will do a better job IMHO.

Cheers,
Rick.
Reply With Quote
  #19  
Old 27-04-2016, 08:32 PM
troypiggo's Avatar
troypiggo (Troy)
Bust Duster

troypiggo is offline
 
Join Date: Nov 2008
Location: Brisbane, Australia
Posts: 4,846
Aah. Gotchya. I do have a KAF-8300, and based on the numbers above, think the good ol' tried and tested L 1x1 and RGB 2x2 will continue to meet my humble needs.
Reply With Quote
  #20  
Old 27-04-2016, 10:14 PM
alpal's Avatar
alpal
Registered User

alpal is offline
 
Join Date: Jan 2012
Location: Melbourne
Posts: 3,602
I've shown this picture before.
Binning can really help to bring out faint detail.
Target: NGC 300.
This was with a KAF8300 & a 10" f4 Newt.

cheers
Allan
Attached Thumbnails
Click for full-size image (Binning comparison NGC 300_3.jpg)
153.1 KB47 views
Reply With Quote
Reply

Bookmarks

Thread Tools
Rate This Thread
Rate This Thread:

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump


All times are GMT +10. The time is now 09:20 PM.

Powered by vBulletin Version 3.8.7 | Copyright ©2000 - 2024, Jelsoft Enterprises Ltd.
Advertisement
Bintel
Advertisement
Testar
Advertisement