Drizzle vs Nondrizzle

[Rodd]

A question that frequently come up-one that I have asked many times.  The only real way to answer it is to have  drizzle and non drizzle versions of the same data side by side.  To remove the effects that different processing can have on the image, processing has been removed from the equation.  No processing has been performed on these images--just calibration, alignment, edge cropping, and gradient removal (dbe).  Even the effects of different instances of DBE were eliminated as the same dbe icon was used for both.  The histogram stretch was carefully performed to as equal as possible.  I find the small stars are much rounder--but any difference can only be seen upon zoom (full resolution).  The big question is at 1:1 and 2:1 (for the non drizzled version) are the images the same--1:1 being different for each image due to the 4x size difference.  Perhaps a  benefit of drizzle is the fact that the image is 4x as big at 1:1--even though if you zoom the non drizzle to 2:1 it is the same.  That and the stars.  The 3rd image is a fully processed stand alone drizzle version.  Its interesting to note that processing the image changed it from 26mb to 16mb. That seems like an awful lot of mega bytes lost to processing.  It should be noted that only light processing was required due to the quantity of data.  If anyone is interested I will post a processed version of the non drizzled image (have to process it).

FSQ 106 with .6x reducer and ASI 1600 with 3um Astrodon Ha:  152 5min subs--almost 13 hours

Non-drizzle

 

Drizzle

 

Fully Processed Drizzle

[tooth_dr]

Lovely images Rodd, and thanks for taking time to review this not widely used processing feature.

[wimvb]

Nice comparison. Did you evaluate and compare the level of noise? In theory, since you use part of the data to restore detail, that data can't be used for noise reduction. A drizzled image should therefore have a slightly higher noise level than a non-drizzled version. It would also be interesting to see how the drizzled image would look compare to the non-drizzled version, after being downsampled at the very end of the process.

[michael8554]

I thought Drizzle was a "Silk Purse from a Pig's Ear" situation, developed by NASA to improve the resolution of the early HST images.

So I suspect your 16Mpix images won't show much improvement?

Michael

[Rodd]

49 minutes ago, tooth_dr said:

Lovely images Rodd, and thanks for taking time to review this not widely used processing feature.

Thanks Doc--Still not sure

Rodd

[Rodd]

2 minutes ago, wimvb said:

Nice comparison. Did you evaluate and compare the level of noise? In theory, since you use part of the data to restore detail, that data can't be used for noise reduction. A drizzled image should therefore have a slightly higher noise level than a non-drizzled version. It would also be interesting to see how the drizzled image would look compare to the non-drizzled version, after being downsampled at the very end of the process.

Yes--the drizzle version has more noise--but not very much more--certainly processing es care of it.  I used SFS to measure noise in the drz and non drz images--but I did not save them directly after integration as I should have.  So--I am reintegrating the drz one now--but 152 subs will take some time--an hour or so.  I will post what I find.  Regarding downsampling--that is what I usually do with my drz'd images and they look the same--but the stars are rounder when you use full resolution mode.  The DRZ images are just too bid and I am not a good enough processor to make them stand up to that level of zoom.  But--very clean mono Ha stacks that have styrong signal--like this one--are a different story.  That's why I used mono Ha--becuase it can actually take a high amount of zoom.  

Rodd

[Rodd]

19 minutes ago, michael8554 said:

I thought Drizzle was a "Silk Purse from a Pig's Ear" situation, developed by NASA to improve the resolution of the early HST images.

So I suspect your 16Mpix images won't show much improvement?

Michael

Well--I am under sampled at 2.46 arcsec/pix--just look at the small stars--square in one and round in the other.  But--I am probably not THAT undersampled because seeing is typically bad, so teh benefit isn't as great as it could bve.

Rodd

[Rodd]

2 hours ago, wimvb said:

Nice comparison. Did you evaluate and compare the level of noise? In theory, since you use part of the data to restore detail, that data can't be used for noise reduction. A drizzled image should therefore have a slightly higher noise level than a non-drizzled version. It would also be interesting to see how the drizzled image would look compare to the non-drizzled version, after being downsampled at the very end of the process.

Nope--the drizzle image has less noise.  If I test them together in SFS, the nondrizzle has about .46 DN and the drizzle about .32.  If I use 2x the pixel scale for the drizzle it is even less (.23).  What does this tell you?

Rodd

[wimvb]

3 hours ago, Rodd said:

What does this tell you?

That one should never trust theory.

Seriously though, downsampling decreased noise, as it should, because you average 4 pixels. But why the drizzled version has lower noise, I don't know. Maybe someone else can explain that.

[MartinB]

Excellent image, can't really make out any difference with drizzle, so I wont be rushing to use it.  Very helpful Rodd, thanks

[Rodd]

4 hours ago, wimvb said:

That one should never trust theory.

Seriously though, downsampling decreased noise, as it should, because you average 4 pixels. But why the drizzled version has lower noise, I don't know. Maybe someone else can explain that.

Well--I din't downsample--I just told SFS that the pixel scale was 4.92 instead of 2.46.  To use 2.46 I would have had to downsample.  I will try this and see what the the numbers say.  

Rodd

[Rodd]

2 hours ago, MartinB said:

Excellent image, can't really make out any difference with drizzle, so I wont be rushing to use it.  Very helpful Rodd, thanks

Its hard to see on the forum because the images at set sizes/  But when you are processing the images and zoom further in than the forum takes you, you can really see a difference in the stars.  The non drizzled version has square small stars and the drizzled version does not.   I have seen demonstrations that show a big difference (other than the stars)  so don't base your decision on my demonstration!  If you use a CCD camera and don't typically collect as many as 50-100 subs per channel (or more), then drizzle is not an option anyway.  

Rodd

[souls33k3r]

Great post @Rodd. Learning so much from you mate

[Rodd]

23 minutes ago, Rodd said:

Well--I din't downsample--I just told SFS that the pixel scale was 4.92 instead of 2.46.  To use 2.46 I would have had to downsample.  I will try this and see what the the numbers say.  

Rodd

Here's a question that I have found very difficult to find an answer to:  Does the size of the dither impact the drizzle quality.  We know that drizzle requires drizzle.  The literature says that you need allot of data with good quality dither.  But it says nothing about dither settings.  Drizzle uses dither movement, so it makes sense that if the movement was large it would have an effect as opposed to it being small.

Rodd

[Rodd]

6 minutes ago, souls33k3r said:

Great post @Rodd. Learning so much from you mate

I hope I don't lead you into the bog where you will be lost for a few thousand years!

Rodd

[souls33k3r]

Just now, Rodd said:

I hope I don't lead you into the bog where you will be lost for a few thousand years!

Rodd

hehehehe you're not and it's quite an interesting topic to follow but i'm going to start a new thread around this any way. Will be posting it in "Image Processing, Help and Techniques" section

[Laurin Dave]

Interesting post and discussion, I took some images of North American and Pelican Nebulas Saturday (posted in Deep sky imaging) using an ASI1600 and a Nikon zoom at 200mm giving a pixel scale of 4" per pixel.  The image was 24x300s and dithered every frame so I had a go at Drizzle integration.  Below is a screen shot of the H555 (Rabbit's Mouth) region with and without drizzle.  Calibrated integrated and screen stretched only. (hopefully the differences will show up). Drizzle image is smoother and stars less blocky.

Dave

 

 

 

 

 

[Rodd]

1 hour ago, Laurin Dave said:

Interesting post and discussion, I took some images of North American and Pelican Nebulas Saturday (posted in Deep sky imaging) using an ASI1600 and a Nikon zoom at 200mm giving a pixel scale of 4" per pixel.  The image was 24x300s and dithered every frame so I had a go at Drizzle integration.  Below is a screen shot of the H555 (Rabbit's Mouth) region with and without drizzle.  Calibrated integrated and screen stretched only. (hopefully the differences will show up). Drizzle image is smoother and stars less blocky.

Your pixel scale was 4.0--mine only 2.46--so you were much more undersampled, so it makes sense that the difference would be more pronounced in yours.  Although I do not see the difference in smoothness-only the stars.  And, to be honest, the roundness comes at the expense of size.  So they are rounder in the drizzle version--but a bit larger.   I guess a bit of MT will restore them to the smaller size.  I see the same in mine--but to a much less extent.  Do you happen to know if the size of the dither affects the drizzle outcome?  Whether ist 1,2,3,4 pixels etc?

Rodd

Edit:  In teh end,, I don't think it was worth the extra 2 hours it took to get the drizzle data.  For me, 152 subs took a long time to drizzle.  

[Laurin Dave]

Hi Rodd,

Yes it is tricky to see/demonstrate the effect on the forum and yes the stars are bigger but in my case they're also not round anyway. I had SGPro on large drizzle, PHD2 scale on 1 so I think its 3 pixels.  Dithering was very rapid, I guess because the mount was very lightly loaded. I'd also guess that the bigger and more random the dither the better the effect would be.

Dave

[Rodd]

10 minutes ago, Laurin Dave said:

Hi Rodd,

Yes it is tricky to see/demonstrate the effect on the forum and yes the stars are bigger but in my case they're also not round anyway. I had SGPro on large drizzle, PHD2 scale on 1 so I think its 3 pixels.  Dithering was very rapid, I guess because the mount was very lightly loaded. I'd also guess that the bigger and more random the dither the better the effect would be.

Dave

I asked this question on the PI forum and the answer was the dither amount didn't matter--but the answer did not come from one of the PI design team, but rather an avid user like us.  So I am not confident in the answer.    We really need to ask one of the dither designers--they would know.

Rodd

 

[wimvb]

4 hours ago, Rodd said:

I asked this question on the PI forum and the answer was the dither amount didn't matter--but the answer did not come from one of the PI design team, but rather an avid user like us.  So I am not confident in the answer.    We really need to ask one of the dither designers--they would know.

Rodd

 

I think it's only the fractional pixel offset of a dither that is used in drizzle integration. Drizzle is based on resampling with a grid that is half or 1/3 the size of the original pixel grid. The stars to be drizzled first have to be moved to the correct location by registering. If a star ends up on exactly the same location of a pixel after registration, you won't be able to drizzle. But because in reality the star ends up slightly different on each frame, drizzle is possible.

There's an old post on this forum about drizzle:

 

Suppose you take 3 seperate images of the dog. If the dogs head ends up exactly on the same pixels every time, or is shifted an exact number of pixels, than resampling with a finer grid won't make a difference. But if the dogs head ends up on slightly different positions (fractions of a pixel) on each image, then each resampled (smaller) new pixel receives slightly different information from each "sub". In this case, the finer grid will show new detail. That's the idea behind drizzle.

Here are a few references. Never mind the math, compare the images in the articles.

https://arxiv.org/pdf/astro-ph/9808087.pdf

https://core.ac.uk/download/pdf/16200502.pdf

If you'd have perfect guiding without dithering, then drizzle also wouldn't work, because the target (star) ends up on the exact same location in every sub. Even hot pixel removal wouldn't work properly. That's why it's important to always dither.

[michael8554]

19 hours ago, Rodd said:

Well--I am under sampled at 2.46 arcsec/pix

Yes Rod, I'd agree that was a tad undersampled, so Drizzle should make a difference.

My bad for only looking at the Megapixels......

Michael

[Rodd]

1 hour ago, wimvb said:

That's why it's important to always dither.

So--dithering 2 pixels or 10 results in the same drizzle sucess,  correct?  the size of the dither is irrelevant as long as there IS dither.  right?

Rodd

[Rodd]

1 hour ago, michael8554 said:

Yes Rod, I'd agree that was a tad undersampled, so Drizzle should make a difference.

My bad for only looking at the Megapixels......

Michael

Then again--your post pointed out that drizzle might not yield as much as one would like if the data is just a bit under sampled.  In fair to poor seeing 2.46 arcsec/pix is probably a bit undersampled, where 4.0 is definitely undersampled.  So it was good.

Rodd

[wimvb]

1 hour ago, Rodd said:

So--dithering 2 pixels or 10 results in the same drizzle sucess,  correct?  the size of the dither is irrelevant as long as there IS dither.  right?

Rodd

That's my understanding of the matter, yes.