Jump to content

  •  

CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.

Photo

Best Time in Workflow to Down Scale, Drizzle?

  • Please log in to reply
29 replies to this topic

#1 h2ologg

h2ologg

    Mariner 2

  • *****
  • topic starter
  • Posts: 276
  • Joined: 04 Jul 2007
  • Loc: Southern California Beach area

Posted 25 August 2019 - 10:47 PM

There are quite a few tutorials on the merits and “how to’s” on drizzling.

 

What I find lacking  is when does one down scale a drizzled image back to the original scale?  I find drizzle labor-intensive/ time consuming on my cpu (i9). Are there benefits to downscaling at the very end of the nonlinear stage or let’s say finishing up deconvolution and denoising in the linear stage of processing?

 

Any good arguments or thoughts on best time in one’s workflow and still gain the benefits.  I’m using PixInsight, solely.  However I don’t think that matters in the process.

 

Thanks,

Dan


  • PirateMike likes this

#2 cfosterstars

cfosterstars

    Mercury-Atlas

  • *****
  • Posts: 2522
  • Joined: 05 Sep 2014
  • Loc: Austin, Texas

Posted 26 August 2019 - 07:22 PM

There are quite a few tutorials on the merits and “how to’s” on drizzling.

 

What I find lacking  is when does one down scale a drizzled image back to the original scale?  I find drizzle labor-intensive/ time consuming on my cpu (i9). Are there benefits to downscaling at the very end of the nonlinear stage or let’s say finishing up deconvolution and denoising in the linear stage of processing?

 

Any good arguments or thoughts on best time in one’s workflow and still gain the benefits.  I’m using PixInsight, solely.  However I don’t think that matters in the process.

 

Thanks,

Dan

The first question to ask is are you undersampled or not. If you are not undersampled and for me - seriously undersampled, then dont drizzle at all. You should do MUREDenoise and Deconvolution. What is your focal length, and pixel scale? I used to drizzle everything by default, but now only do it on my SV80ST2 with my ASI071MC-PRO. For my other setups I no longer do drizzle. 


  • h2ologg and MattC867 like this

#3 drmikevt

drmikevt

    Viking 1

  • *****
  • Posts: 848
  • Joined: 22 Jun 2015
  • Loc: Burlington, VT

Posted 26 August 2019 - 08:01 PM

I don't know... I've heard the 'only drizzle under sampled data' rule before.  And it makes sense.  But, I can unequivocally say that, at .62 arcsec/pix, I can see clear improvements to star shapes, contrast and other small details after drizzle.  I thought I would stop using drizzle when moving from my SV80 to my current RC10, but I was surprised at how much improvement I could see - If I look close, to be sure, but aren't we all pixel peepers here anyway?

 

And to address the OP - I don't ever both downscaling so I can't help (sorry). 


Edited by drmikevt, 26 August 2019 - 08:06 PM.

  • elmiko likes this

#4 h2ologg

h2ologg

    Mariner 2

  • *****
  • topic starter
  • Posts: 276
  • Joined: 04 Jul 2007
  • Loc: Southern California Beach area

Posted 26 August 2019 - 08:42 PM

I’m running a CFF 105 F6 oiled apo on a 10Micron 1000hps Mount unguided.  Using the ZWO ASI183mmPro and an image scale of 0.79”/pixel. Undersampled for exceptional seeing in my dark sky cabin June Lake California.  

 

I’m processing 145 x 60 s of LRGB summertime data of M33 (as well as some other stuff). Data looks great.

 

Still looking for best practice for down scaling back to original.  Will I see continued benefits from drizzled image processing after deconvolution and NR in the linear stage?  Or no real benefit when stretching, nonlinear noise reduction and sharpening?  Again, best practice for even the slightest amount of image processing improvement. 

 

Thanks for your thoughts.

Dan



#5 cfosterstars

cfosterstars

    Mercury-Atlas

  • *****
  • Posts: 2522
  • Joined: 05 Sep 2014
  • Loc: Austin, Texas

Posted 26 August 2019 - 10:54 PM

I’m running a CFF 105 F6 oiled apo on a 10Micron 1000hps Mount unguided.  Using the ZWO ASI183mmPro and an image scale of 0.79”/pixel. Undersampled for exceptional seeing in my dark sky cabin June Lake California.  

 

I’m processing 145 x 60 s of LRGB summertime data of M33 (as well as some other stuff). Data looks great.

 

Still looking for best practice for down scaling back to original.  Will I see continued benefits from drizzled image processing after deconvolution and NR in the linear stage?  Or no real benefit when stretching, nonlinear noise reduction and sharpening?  Again, best practice for even the slightest amount of image processing improvement. 

 

Thanks for your thoughts.

Dan

You should just try both ways and see which is better on your data. I have seen a HUGE benefit in my data, but my equipment is significantly lower grade than yours and my skies are Bortle 5. I have seen great results with my Refractor at 1.19"/pixel not using drizzle and I never use it with my SCT at 0.36"/pixel.



#6 WadeH237

WadeH237

    Aurora

  • *****
  • Posts: 4682
  • Joined: 24 Feb 2007
  • Loc: Snohomish, WA

Posted 27 August 2019 - 07:49 AM

I don't know... I've heard the 'only drizzle under sampled data' rule before.  And it makes sense.  But, I can unequivocally say that, at .62 arcsec/pix, I can see clear improvements to star shapes, contrast and other small details after drizzle.  I thought I would stop using drizzle when moving from my SV80 to my current RC10, but I was surprised at how much improvement I could see - If I look close, to be sure, but aren't we all pixel peepers here anyway?

 

And to address the OP - I don't ever both downscaling so I can't help (sorry). 

Drizzle is a blurring algorithm.

 

If you see improvements afterwards, even with over sampled data, then it sounds like you are seeing some denoising benefits.  I agree with cfosterstars about MureDenoise.  Used correctly, I suspect that it would give you a great result.  Now that I've got it figured out, I can't imaging not using it.



#7 spokeshave

spokeshave

    Vanguard

  • *****
  • Posts: 2028
  • Joined: 08 Apr 2015

Posted 27 August 2019 - 10:26 AM

Drizzle is a blurring algorithm.

I'm not sure how drizzle can be viewed as a blurring algorithm. While it isn't exactly a sharpening algorithm either, it does recover information lost to sampling and the net effect can be a more detailed final image. The amount of detail that can be recovered is directly related to how undersampled the image scale is. 

 

Tim


Edited by spokeshave, 27 August 2019 - 10:26 AM.

  • elmiko and Jon Rista like this

#8 cfosterstars

cfosterstars

    Mercury-Atlas

  • *****
  • Posts: 2522
  • Joined: 05 Sep 2014
  • Loc: Austin, Texas

Posted 27 August 2019 - 10:42 AM

I'm not sure how drizzle can be viewed as a blurring algorithm. While it isn't exactly a sharpening algorithm either, it does recover information lost to sampling and the net effect can be a more detailed final image. The amount of detail that can be recovered is directly related to how undersampled the image scale is. 

 

Tim

Tim,

 

I agree with you. As I stated above, I used to Drizzle most of the my images from my refractors at 1.19 and 2.02"/pixel image scale. For the shorter focal length, I still drizzle since it is the most undersampled and the widest field. My APOs have been down waiting for a part for my new focuser so I have been mostly imaging with my SCT at 2159mm F/L. For my SCT at 0.36"/pixel, Drizzle make no sense at all and this is why I started to do MUREDenoise. I spend a bunch of time figuring out how to make it work and now its so good, that I dont want to NOT use it. When you do MUREDenoise and a good DBE on the master frames first, I fine the Deconvolution is both easier and significantly more effective with much few artifacts. This make sense because of the different length scales that Deconvolution and MUREDenoise work at. Two of the big issues with deconvolution are

 

1) that you have to be careful not to over do it since you sharpen the noise too much and get ugly background even if you protect it - especially in the no mans land between the bright and background portions of the image. Since MURE Denoise kill the noise across the image and not just in the background, you can get more and better quality sharpening without the ugly artifacts.

 

2) it can take many iterations to find the right global dark and global bright setting for deconvolution. What I have found is this is much easier and I end up using that same setting for just about all images. I test it to be sure, but it just seems to work. 

 

I have gone back to my APO data at 1.19"/pixel. It is marginally undersampled. You can see improvement with drizzle for sure. But I have found that the combination of MUREDenoise and a good deconvolution leads to a significantly better result overall. I have even now gone back to old data sets and regretted that I did not keep the non-drizzled masters. 

 

This is just my opinion and experience, but I highly suggest that you try it. MUREDenoise takes a bit of effort to learn to use, but much less that Deconvolution took.


  • h2ologg likes this

#9 spokeshave

spokeshave

    Vanguard

  • *****
  • Posts: 2028
  • Joined: 08 Apr 2015

Posted 27 August 2019 - 12:54 PM

I use MUREDenoise where I can. With the QHY163 on the Tak, my image scale is right at 1.5"/px and drizzle really helps. Since I tend to collect hundreds of subs, I can use a fairly large drop shrink (typically 0.5) which improves the effectiveness of drizzle. Unfortunately, I can't use MUREDenoise on drizzled images, but I also tend to get fairly deep data anyway, so my noise reduction needs are modest. 

 

Deconvolution works very well on drizzled images since the PSF is larger (in pixels) and much more easily characterized. Protecting the background can be challenging, but there are very good techniques for that.

 

Tim



#10 WadeH237

WadeH237

    Aurora

  • *****
  • Posts: 4682
  • Joined: 24 Feb 2007
  • Loc: Snohomish, WA

Posted 27 August 2019 - 12:54 PM

I'm not sure how drizzle can be viewed as a blurring algorithm. While it isn't exactly a sharpening algorithm either, it does recover information lost to sampling and the net effect can be a more detailed final image. The amount of detail that can be recovered is directly related to how undersampled the image scale is. 

 

Tim

Drizzle works by taking light from a single pixel and distributing it among the neighboring pixels according to an algorithm.  Process the same data with and without Drizzle and see which one has the better SNR.


Edited by WadeH237, 27 August 2019 - 12:54 PM.


#11 spokeshave

spokeshave

    Vanguard

  • *****
  • Posts: 2028
  • Joined: 08 Apr 2015

Posted 27 August 2019 - 02:08 PM

Drizzle works by taking light from a single pixel and distributing it among the neighboring pixels according to an algorithm.  Process the same data with and without Drizzle and see which one has the better SNR.

There is very little loss in SNR from drizzling. The whole concept of drizzling was developed for the Hubble Wide Field Camera to recover information lost to undersampling without a significant penalty to SNR. 

 

Tim


  • Jon Rista, PhotonHunter1 and 2ghouls like this

#12 Jon Rista

Jon Rista

    ISS

  • *****
  • Posts: 23548
  • Joined: 10 Jan 2014
  • Loc: Colorado

Posted 27 August 2019 - 03:56 PM

Drizzle works by taking light from a single pixel and distributing it among the neighboring pixels according to an algorithm.  Process the same data with and without Drizzle and see which one has the better SNR.

The algorithm involves a "shrinking" of the relative size of the original pixel before the information is distributed. As such, it has the potential to improve resolution, at the very least. Once you drizzle enough frames, the SNR doesn't really change...it isn't really worse, nor really better, especially not when the images are normalized. With drizzling, you do need "enough" frames, and exactly how many that will be depends on exactly how much shrink you apply to the source pixels. Without enough data, then SNR might suffer a bit with the drizzle...however, that is only at native resolution. Normalize it (i.e. downsample the drizzle back to native size), and there really is no loss in SNR. It IS the same information, after all.

 

I would say though that drizzling is overall a better algorithm for integration than your normal pixel stacking. I often have funky star shapes with normal integration that disappear with drizzling, and when I use a small enough drop shrink setting in PI, there is a slight improvement in resolution as well (which is then easier to improve with deconvolution with the better-sampled information of the drizzled image). 


  • psandelle, elmiko, cfosterstars and 1 other like this

#13 Jerry Lodriguss

Jerry Lodriguss

    Vendor

  • *****
  • Vendors
  • Posts: 7225
  • Joined: 19 Jul 2008
  • Loc: Voorhees, NJ

Posted 27 August 2019 - 04:57 PM

Still looking for best practice for down scaling back to original.  Will I see continued benefits from drizzled image processing after deconvolution and NR in the linear stage?  Or no real benefit when stretching, nonlinear noise reduction and sharpening?  Again, best practice for even the slightest amount of image processing improvement. 

Hi Dan,

 

Drizzle improves spatial resolution.

Downsampling reduces apparent noise.

 

But the cost of the noise reduction by downsampling is loss of spatial resolution. So always save your original high-res drizzled image, it has the most information in it. Any downsampling after should be based on output use. 

 

So,

  • Drizzle first
  • Decon
  • Stretch
  • Downsample
  • Additional noise reduction if necessary and additional sharpening if necessary

 

Generally, it is best practice to downsample to the exact resolution you need for output (screen, print, press) then examine the image and see if it needs any further noise reduction.

 

Sharpening also should be based on output use after downsamping. For example, you need more sharpening when you print than just for display on a monitor.

 

Also, you can experiment with different downsampling algorithms to see which you like the best.

 

Jerry


  • h2ologg and psandelle like this

#14 freestar8n

freestar8n

    Vendor - MetaGuide

  • *****
  • Vendors
  • Posts: 8737
  • Joined: 12 Oct 2007

Posted 27 August 2019 - 06:29 PM

If you drizzle using a sub pixel window then each drizzle pixel will contribute its data to fewer output pixels than it would have otherwise. By not touching as many pixels there is definitely reduced snr and there may also be increased detail.

I use fairly small pixels so I never use it.

Professional work usually involves very precise dither patterns rather than random so they definitely get a resolution win. But again less signal contributes to each pixel so a loss of snr.

Frank

#15 Monkeybird747

Monkeybird747

    Viking 1

  • *****
  • Posts: 546
  • Joined: 31 Aug 2017
  • Loc: Louisville, KY

Posted 28 August 2019 - 12:40 AM

Still looking for best practice for down scaling back to original.  Will I see continued benefits from drizzled image processing after deconvolution and NR in the linear stage?  Or no real benefit when stretching, nonlinear noise reduction and sharpening?  Again, best practice for even the slightest amount of image processing improvement.

I do the same as Jerry. Normal integration, drizzle integration, initial MLT noise reduction, Deconvolution, stretch, resample 50%, and usually ACDNR next.



#16 h2ologg

h2ologg

    Mariner 2

  • *****
  • topic starter
  • Posts: 276
  • Joined: 04 Jul 2007
  • Loc: Southern California Beach area

Posted 28 August 2019 - 07:47 PM

Thanks all!

Learned something new in the question, both in terminology and in processing. I’ve avoided the “MureDenoise” script and will have to experiment.

 

So “normalize = downsample”.

 

Normalize after stretch, then sharpen to taste.

Best!

Dan



#17 Jon Rista

Jon Rista

    ISS

  • *****
  • Posts: 23548
  • Joined: 10 Jan 2014
  • Loc: Colorado

Posted 28 August 2019 - 08:21 PM

Thanks all!

Learned something new in the question, both in terminology and in processing. I’ve avoided the “MureDenoise” script and will have to experiment.

 

So “normalize = downsample”.

 

Normalize after stretch, then sharpen to taste.

Best!

Dan

Normalize is not necessarily downsampling. Normalization, in this context, means to make the extraneous aspects of each image in a comparison the same. So, you might downsample one image to the same size as another, or upsample an image to the size of another. Normalization doesn't necessarily require downsampling. And, in fact, sometimes you learn more if you compare both ways, by downsampling one to the other, then upsampling the other to the one, to see what the differences of each approach are. 

 

Normalization aims to put everything on level ground in the same playing field, reduce the number of variables to as few as possible (or zero, ideally.) 



#18 h2ologg

h2ologg

    Mariner 2

  • *****
  • topic starter
  • Posts: 276
  • Joined: 04 Jul 2007
  • Loc: Southern California Beach area

Posted 28 August 2019 - 08:55 PM

Normalize is not necessarily downsampling. Normalization, in this context, means to make the extraneous aspects of each image in a comparison the same. So, you might downsample one image to the same size as another, or upsample an image to the size of another. Normalization doesn't necessarily require downsampling. And, in fact, sometimes you learn more if you compare both ways, by downsampling one to the other, then upsampling the other to the one, to see what the differences of each approach are. 

 

Normalization aims to put everything on level ground in the same playing field, reduce the number of variables to as few as possible (or zero, ideally.) 

Thanks Jon for clarifying.

Dan



#19 PhotonHunter1

PhotonHunter1

    Viking 1

  • *****
  • Posts: 597
  • Joined: 19 Mar 2015
  • Loc: Northwest IL (for now)

Posted 09 September 2019 - 11:06 PM

Hi Dan,

 

Drizzle improves spatial resolution.

Downsampling reduces apparent noise.

 

But the cost of the noise reduction by downsampling is loss of spatial resolution. So always save your original high-res drizzled image, it has the most information in it. Any downsampling after should be based on output use. 

 

So,

  • Drizzle first
  • Decon
  • Stretch
  • Downsample
  • Additional noise reduction if necessary and additional sharpening if necessary

 

Generally, it is best practice to downsample to the exact resolution you need for output (screen, print, press) then examine the image and see if it needs any further noise reduction.

 

Sharpening also should be based on output use after downsamping. For example, you need more sharpening when you print than just for display on a monitor.

 

Also, you can experiment with different downsampling algorithms to see which you like the best.

 

Jerry

I’ve been following this thread with a lot of interest - and as a result have used drizzle integration for the first time. I’m very happy with the results but unclear how to downsample. If I’m following you correctly, I would go through ALL my post processing and non-linear steps (even creating starless narrowband images) and then downsample my final image - applying any NR and sharpening if necessary. Is there a program in PI that you use to downsample drizzled images?



#20 2ghouls

2ghouls

    Viking 1

  • *****
  • Posts: 743
  • Joined: 19 Sep 2016

Posted 09 September 2019 - 11:42 PM

Is there a program in PI that you use to downsample drizzled images?


The process is called “resample” in PI. You can use a percentage. If you want to go back to the original resolution and you did 2x drizzle, you would resample at 50%. If you have time, I would suggest trying the different interpolation algorithms to find the best one for your data, as the one it picks automatically is not always best in my experience.

Edited by 2ghouls, 09 September 2019 - 11:44 PM.

  • h2ologg, psandelle, Jon Rista and 1 other like this

#21 PhotonHunter1

PhotonHunter1

    Viking 1

  • *****
  • Posts: 597
  • Joined: 19 Mar 2015
  • Loc: Northwest IL (for now)

Posted 11 September 2019 - 04:07 PM

The process is called “resample” in PI. You can use a percentage. If you want to go back to the original resolution and you did 2x drizzle, you would resample at 50%. If you have time, I would suggest trying the different interpolation algorithms to find the best one for your data, as the one it picks automatically is not always best in my experience.

Perfect. Just the information I was hoping for. Thank you!



#22 h2ologg

h2ologg

    Mariner 2

  • *****
  • topic starter
  • Posts: 276
  • Joined: 04 Jul 2007
  • Loc: Southern California Beach area

Posted 12 September 2019 - 07:27 PM

Wow! Influenced by the initial posts, I deviated from the drizzle workflow and used PI script MureDenoise instead (since this script and drizzle aren’t compatible).  I may never go back to drizzle, save for planetary imaging.  My data has never been so “clean” in the early stages of mono processing.  I guess if I’m looking for resolution I’ll be drizzling and if noise reduction is of importance MureDenoise will be in my early workflow.

 

Dan



#23 Monkeybird747

Monkeybird747

    Viking 1

  • *****
  • Posts: 546
  • Joined: 31 Aug 2017
  • Loc: Louisville, KY

Posted 12 September 2019 - 07:44 PM

Well, don’t keep us in suspense 😉. Let’s see the data! I’ve never tried that script.

#24 h2ologg

h2ologg

    Mariner 2

  • *****
  • topic starter
  • Posts: 276
  • Joined: 04 Jul 2007
  • Loc: Southern California Beach area

Posted 12 September 2019 - 09:27 PM

As requested...

MureDenoiseExample.jpg

Not sure the jpeg does the background justice with its 85% reduction in quality and its inherent lossy format.  Red channel is my worst.  At the June Lake California dark site, I pick up IFN often (my M15 work to date is inundated using an OSC) especially the red.  I've included an early RGB trial prior to any post linear cleanup. It's looking promising, and in fact the Lum channel has yet to be added!

Dan


Edited by h2ologg, 12 September 2019 - 09:28 PM.

  • NorthField likes this

#25 Jon Rista

Jon Rista

    ISS

  • *****
  • Posts: 23548
  • Joined: 10 Jan 2014
  • Loc: Colorado

Posted 12 September 2019 - 09:50 PM

As requested...

attachicon.gif MureDenoiseExample.jpg

Not sure the jpeg does the background justice with its 85% reduction in quality and its inherent lossy format.  Red channel is my worst.  At the June Lake California dark site, I pick up IFN often (my M15 work to date is inundated using an OSC) especially the red.  I've included an early RGB trial prior to any post linear cleanup. It's looking promising, and in fact the Lum channel has yet to be added!

Dan

Upload full size, uncompressed (or 100% JPEG quality) versions somewhere, and put the links here.




CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.


Recent Topics






Cloudy Nights LLC
Cloudy Nights Sponsor: Astronomics