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

Drizzle - Example thread

  • Please log in to reply
38 replies to this topic

#1 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 18 July 2019 - 11:00 PM

I was comparing a regular integration and 2x drizzled integration tonight, and thought others might be interested to see how great it can work on even slightly undersampled data. This is at:

  • 1.67" per pixel
  • 86 subexposures
  • 2x drizzle with 0.9 drop shrink (PixInsight defaults)

 

driz1.jpg

 

driz2.jpg

 

Would be interested to see others results at various pixel scales.

 

Cheers, Nico


  • Stelios, H-Alfa, psandelle and 17 others like this

#2 jerahian

jerahian

    Messenger

  • *****
  • Posts: 495
  • Joined: 02 Aug 2018
  • Loc: Maine

Posted 19 July 2019 - 12:28 AM

Isn't that a beaut!?  What great technology we have at our disposal...

 

Thanks for sharing this test with us!

 

Ara



#3 JP50515

JP50515

    Viking 1

  • *****
  • Posts: 584
  • Joined: 23 Feb 2017

Posted 19 July 2019 - 01:29 AM

Drizzling is easily one of the things I understand least in this hobby. Don't know how/when to use it or how it works...but it's cool. 



#4 james7ca

james7ca

    Fly Me to the Moon

  • *****
  • Posts: 7188
  • Joined: 21 May 2011
  • Loc: San Diego, CA

Posted 19 July 2019 - 04:59 AM

How did you resize those images and did the non-drizzled sample start at the same scale as the original capture?

 

Also, what was your camera pixel size and the aperture and focal length of the scope?

 

Below are two links that compare drizzled to non-drizzled results when using short-focus camera lenses:

 

  https://www.cloudyni...-2#entry9150677

 

  https://www.cloudyni...dpost&p=6956474



#5 Der_Pit

Der_Pit

    Viking 1

  • -----
  • Posts: 726
  • Joined: 07 Jul 2018
  • Loc: La Palma

Posted 19 July 2019 - 06:48 AM

OK, not really a drizzle example, only a historic comparison:

 

m82_2.jpg  m82_1.jpg

 

This is from a small article I wrote for German magazine 'Sterne und Weltraum' in 1993.

It compares the result of a simple subpixel shift addition (left) with the sum of the same images, just upscaled (4× linear interpolation) before processing.  It was inspired by (what I think was) the 'initial idea for drizzle', a paper from Hook and Lucy in 1993 (Co-adding images with different PSFs).

 

Data had been acquired with an ST-4 attached to a 200mm F/4.5 Newton camera. 

laugh.gif


Edited by Der_Pit, 19 July 2019 - 06:49 AM.

  • bobzeq25, 2ghouls and RossW like this

#6 mewmartigan

mewmartigan

    Viking 1

  • *****
  • Posts: 909
  • Joined: 02 Jul 2008
  • Loc: outside of Pittsburgh, PA

Posted 19 July 2019 - 07:41 AM

When using Pixinsight does Drizzle have to be done manually? Or is it now automated as part of the batch pre processing script?



#7 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 19 July 2019 - 08:31 AM

How did you resize those images and did the non-drizzled sample start at the same scale as the original capture?

 

Also, what was your camera pixel size and the aperture and focal length of the scope?

 

Below are two links that compare drizzled to non-drizzled results when using short-focus camera lenses:

 

  https://www.cloudyni...-2#entry9150677

 

  https://www.cloudyni...dpost&p=6956474

1. Yes, the non-drizzled sample is at the same scale as the original capture.

I didn't exactly resize the images for this illustration, I just zoomed in on PI windows to get the same field and took screenshots. I'd be happy to try another test where I drizzle and then resample it down to original capture size. From experience, it will look roughly the same as what I've shown here.

 

2. 3.75 μm, 86mm, 464mm. 

 

3. Thanks for the links. I'll add to your image quality thread shortly. Hadn't seen it before.



#8 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 19 July 2019 - 08:33 AM

When using Pixinsight does Drizzle have to be done manually? Or is it now automated as part of the batch pre processing script?

You would still have to run DrizzleIntegration separately after you run BPP, but BPP does have an option to generate the drizzle data. So I'd call it semi-automated (one extra step needed at the end).

 

Screen Shot 2019-07-19 at 9.25.02 AM.png



#9 jerahian

jerahian

    Messenger

  • *****
  • Posts: 495
  • Joined: 02 Aug 2018
  • Loc: Maine

Posted 19 July 2019 - 08:34 AM

When using Pixinsight does Drizzle have to be done manually? Or is it now automated as part of the batch pre processing script?


No, it will generate the drizzle data, but you will need to run DrizzleIntegration manually after. You should do the integration part manually always anyway, so you can test different parameters on the integration step without having to go through the whole BPP setup again. I just use BPP to generate my calibrated files.

-Ara

#10 mic1970

mic1970

    Viking 1

  • -----
  • Posts: 662
  • Joined: 09 Aug 2018
  • Loc: St. Louis Metro Area

Posted 19 July 2019 - 08:36 AM

popcorn.gif



#11 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 19 July 2019 - 08:45 AM

Drizzling is easily one of the things I understand least in this hobby. Don't know how/when to use it or how it works...but it's cool. 

Here is a quick summary.

 

History: The algorithm was developed for the images coming back from Hubble Deep Field, which were undersampled. The original paper is actually quite readable and explains the theory behind drizzle very well: http://www.stsci.edu...er/drizzle.html

 

When to use it: You use it when you have undersampled, well-dithered data.  If you look at the paper, you will see why it only works on dithered data. To properly fill the finer grid, I would also suggest you only use drizzle when you have at least 25 subs, and the more the better.

 

How to use it: In PixInsight, when you register your subs, turn on the "generate drizzle data" option. Then you integrate normally, but add your drizzle data with the button "Add Drizzle data". Finally, you load the drizzle data in to the DrizzleIntegration process and run it. I typically use the default parameters.

 

Why to use it: Main reason is better looking star profiles and some improvement to fine details. It will also increase your resolution, but I usually resample back down (at least for display on the web). 


  • RazvanUnderStars and JP50515 like this

#12 descott12

descott12

    Viking 1

  • -----
  • Posts: 880
  • Joined: 28 Sep 2018
  • Loc: Charlotte, NC

Posted 19 July 2019 - 08:45 AM

Fantastic example.



#13 james7ca

james7ca

    Fly Me to the Moon

  • *****
  • Posts: 7188
  • Joined: 21 May 2011
  • Loc: San Diego, CA

Posted 19 July 2019 - 09:12 AM

However, drizzle will decrease your signal-to-noise (somewhat) although some think that the drizzled noise is somewhat easier on the eyes.

 

Also, it kind of looks like the non-drizzled images were resampled with the nearest neighbor algorithm which is going to produce pretty bad looking results. Perhaps a more practical comparison would be to resize using a better algorithm (but I'm still wondering how the OP prepared these examples, see post #4).

Attached Thumbnails

  • Resampled Images.jpg

Edited by james7ca, 19 July 2019 - 09:23 AM.

  • Jon Rista likes this

#14 johnsoda

johnsoda

    Viking 1

  • *****
  • Moderators
  • Posts: 952
  • Joined: 30 May 2014
  • Loc: SW Ohio, USA

Posted 19 July 2019 - 11:06 AM

Another example where drizzling worked well.

 

https://www.cloudyni...orks/?p=8908428


  • 2ghouls likes this

#15 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 19 July 2019 - 11:48 AM

However, drizzle will decrease your signal-to-noise (somewhat) although some think that the drizzled noise is somewhat easier on the eyes.

 

Also, it kind of looks like the non-drizzled images were resampled with the nearest neighbor algorithm which is going to produce pretty bad looking results. Perhaps a more practical comparison would be to resize using a better algorithm (but I'm still wondering how the OP prepared these examples, see post #4).

Interesting. See post #7 for how I prepared these examples.

 

To be clear, you are saying a fairer test would be to: resample the non-drizzled data 2x bigger with the Lanczos-4 algorithm, and then take the crops?



#16 james7ca

james7ca

    Fly Me to the Moon

  • *****
  • Posts: 7188
  • Joined: 21 May 2011
  • Loc: San Diego, CA

Posted 19 July 2019 - 11:56 AM

...I didn't exactly resize the images for this illustration, I just zoomed in on PI windows to get the same field and took screenshots. I'd be happy to try another test where I drizzle and then resample it down to original capture size...

Okay, then the zoom and screenshot could be an "issue." That's going to use a very low quality resampling and you really shouldn't try to compare the images that way.

 

If you want to compare the results between the non-drizzled but resized ORIGINAL and the 2X drizzled output then you need to do an actual resample operation on the original image. What I'd suggest is that you start with a verified, same-scale copy of the original image and then use PixInsight's Resample tool (under Process->Geometry->Resample) to do a 200% rescale on the original image (just use the default or automatic rescale method). Then compare that output to the 2X drizzled image (both at full size or 1X). I think you'll find that there is less difference between the images than your original screen shots indicate.

 

Then, if you want to do a 2X blowup of the drizzled image use the Resample tool on that also and compare that to a 4X resample of the original image.

 

Below is an example of how the simple on-screen zoom looks in comparison to an image that has actually been resampled and saved as a new file. The slight difference in the brightness of the images is from the color management that was done by Photoshop on the screenshot that was captured by Mac OS. Both of these images are from the same non-drizzled image that has either been resampled by 400% in PixInsight and saved as a new image or simply zoomed by 4X in PixInsight and then copied with a screenshot.

Attached Thumbnails

  • Resample Versus Simple Zoom.jpg

Edited by james7ca, 19 July 2019 - 12:43 PM.

  • 2ghouls likes this

#17 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 19 July 2019 - 12:27 PM

Okay, then the zoom and screenshot could be an "issue." That's going to use a very low quality resampling and you really shouldn't try to compare the images that way.

If you want to compare the results between the non-drizzled but resized ORIGINAL and the 2X drizzled output then you need to do an actual resample operation on the original image. What I'd suggest is that you start with a verified, same-scale copy of the original image and then use PixInsight's Resample tool (under Process->Geometry->Resample) to do a 200% rescale on the original image (just use the default or automatic rescale method). Then compare that output to the 2X drizzled image (both at full size or 1X). I think you'll find that there is less difference between the images than your original screen shots indicate.

Thanks James. Here is the result of doing as you suggest. My analysis is the small stars do look better on the non-drizzled data in this comparison than in my original 'zoom' example, but the medium size stars have taken on an unfortunate diamond shape from the resampling at 200%. 

 

test2.jpg

 

test1.jpg



#18 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 19 July 2019 - 12:39 PM

One more example:

 

Screen Shot 2019-07-19 at 1.37.47 PM.png



#19 james7ca

james7ca

    Fly Me to the Moon

  • *****
  • Posts: 7188
  • Joined: 21 May 2011
  • Loc: San Diego, CA

Posted 19 July 2019 - 12:40 PM

Yes, but what you are seeing there is the advantage of doing the drizzle. If you look closely at undersampled stars they will often have a square or diamond shape and if you resample the original that defect just becomes more apparent (along with possible resample artifacts, you can't get something for nothing). However, the drizzle will produce rounder looking stars.

 

That said, you might get a better looking resample with a different algorithm and I assume that you just used the automatic interpolation algorithm (which is still better than the zoom and screen grab you posted earlier). The only reason why I used Lanczos-4 in my earlier example is because I wanted to know what algorithm I was using so I could include that info in my post, but a different algorithm may produce an even better looking resample.



#20 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 19 July 2019 - 12:49 PM

That said, you might get a better looking resample with a different algorithm and I assume that you just used the automatic interpolation algorithm (which is still better than the zoom and screen grab you posted earlier). The only reason why I used Lanczos-4 in my earlier example is because I wanted to know what algorithm I was using so I could include that info in my post, but a different algorithm may produce an even better looking resample.

Oh, sorry, I should have said: I tried automatic and then Lanczos-4, and the Lanczos-4 looked slightly better so that's what you are seeing here in these new examples. I don't know if I have the patience to try every algorithm, but I see your point that this is a better way to compare the effect of drizzling versus the zooming method I used originally. 



#21 james7ca

james7ca

    Fly Me to the Moon

  • *****
  • Posts: 7188
  • Joined: 21 May 2011
  • Loc: San Diego, CA

Posted 19 July 2019 - 01:31 PM

I think someone here on CN did a series on resampling results (different algorithms) in PixInsight. It may have been Jon Rista and it could have been several years ago and I can't remember whether it covered both scaling up and down. However, you probably can't go strictly by someone else's examples since things like the original sampling, the sensor, and the optics could change the results. Also, even changes in the subject may require different resampling (when you get down to looking really close).

 

Frankly, however, I almost never scale up my images although drizzling can even make the 1X (original) scaling look better (after you resample the drizzled image back down to the original size). Anyway, based upon your previous comments it seems that you often do the latter (rescale the drizzled image back to the original size).



#22 Jon Rista

Jon Rista

    ISS

  • *****
  • Posts: 23352
  • Joined: 10 Jan 2014
  • Loc: Colorado

Posted 19 July 2019 - 02:48 PM

I think someone here on CN did a series on resampling results (different algorithms) in PixInsight. It may have been Jon Rista and it could have been several years ago and I can't remember whether it covered both scaling up and down. However, you probably can't go strictly by someone else's examples since things like the original sampling, the sensor, and the optics could change the results. Also, even changes in the subject may require different resampling (when you get down to looking really close).

 

Frankly, however, I almost never scale up my images although drizzling can even make the 1X (original) scaling look better (after you resample the drizzled image back down to the original size). Anyway, based upon your previous comments it seems that you often do the latter (rescale the drizzled image back to the original size).

I think I did a comparison at one point... It was quite a while ago, now, though...and I'm not finding any threads that seem to be relevant... 

 

In my experience, drizzling's benefits are in a more pleasing noise profile, and rounder stars, than simply resampling. I, too, have had problems with diamond shaped stars (even without resampling), but with drizzling they usually end up nice and round. 

I usually use drizzling "temporarily", in that I'll drizzle, do deconvolution in the better-sampled drizzled data, maybe do some noise reduction and sharpening, then downsample back to normal size again and continue processing. Otherwise, you have to deal with a LOT of pixels throughout your entire post-processing workflow... That could be 60, 80, 100 megapixels, possibly with three channels. That much data takes a lot longer to process...



#23 BBC

BBC

    Sputnik

  • -----
  • Posts: 40
  • Joined: 07 Jun 2014
  • Loc: North Carolina, USA

Posted 19 July 2019 - 03:41 PM

Interesting thread. I’ve always had drizzle integration a standard part of my integration workflow.

Curious. Is drizzle integration at a scale of 1 more or less beneficial than scale of 2 and then resampling back down or do they just end up the same?


Sent from my iPhone using Tapatalk Pro

Edited by BBC, 19 July 2019 - 03:43 PM.


#24 2ghouls

2ghouls

    Viking 1

  • *****
  • topic starter
  • Posts: 699
  • Joined: 19 Sep 2016

Posted 20 July 2019 - 01:22 PM

I almost never scale up my images although drizzling can even make the 1X (original) scaling look better (after you resample the drizzled image back down to the original size). Anyway, based upon your previous comments it seems that you often do the latter (rescale the drizzled image back to the original size).I usually use drizzling "temporarily", in that I'll drizzle, do deconvolution in the better-sampled drizzled data, maybe do some noise reduction and sharpening, then downsample back to normal size again and continue processing. 

I usually use drizzling "temporarily", in that I'll drizzle, do deconvolution in the better-sampled drizzled data, maybe do some noise reduction and sharpening, then downsample back to normal size again and continue processing. 

Curious. Is drizzle integration at a scale of 1 more or less beneficial than scale of 2 and then resampling back down or do they just end up the same?

I've always drizzled 2x, done some processing, and then resampled back down (what Jon describes as drizzling "temporarily" above). Comparing that to not drizzling at all, I think my end result looks better. But BBC brings up a good point here, what about using DrizzleIntegration at 1x? I searched on CN, and came across this thread from you Jon about doing just that. But it sounds like you are now drizzling 2x and then resampling. Any reason you no longer drizzle at 1x? Also, of interest to me is how that thread started comparing drizzle to MureDenoise script. I've tried using MureDenoise, and have never been able to good results from it with my ASI1600. Maybe time for another thread.



#25 Jon Rista

Jon Rista

    ISS

  • *****
  • Posts: 23352
  • Joined: 10 Jan 2014
  • Loc: Colorado

Posted 20 July 2019 - 01:41 PM

I've always drizzled 2x, done some processing, and then resampled back down (what Jon describes as drizzling "temporarily" above). Comparing that to not drizzling at all, I think my end result looks better. But BBC brings up a good point here, what about using DrizzleIntegration at 1x? I searched on CN, and came across this thread from you Jon about doing just that. But it sounds like you are now drizzling 2x and then resampling. Any reason you no longer drizzle at 1x? Also, of interest to me is how that thread started comparing drizzle to MureDenoise script. I've tried using MureDenoise, and have never been able to good results from it with my ASI1600. Maybe time for another thread.

Drizzling integrates the data in a different way than normal integration. The way it distributes input information into the output image is, IMO, better and more refined. So, even if  you drizzle 1x, here is still usually an improvement, such as no diamond stars, over regular integration.

 

Drizzling 1x is not quite the same as drizzling 2x, because there is no increase in sampling rate. If you are not well sampled in your source data, then drizzling 1x will not offer any benfits to say deconvolution like drizzling 2x will, since deconvolution benefits from well sampled data (better PSF modeling, more information to work with when redistributing signal). 


  • RazvanUnderStars, 2ghouls and JP50515 like this


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