Harnessing Dither, Drizzle, and Deconvolution for Greater Detail and Resolution

John Hayes:

Nick Grundy:
Steeve, I love this approach. It's definitely worth the time and thanks for sharing it with us. 

I do remember Russ mentioned at NEAIC that BXT does best on undersampled data. I assume that's pretty much what you accomplish with the 2x drizzle, no? (I could be confusing this aspect)

2x drizzle over-samples the image.

Dithering typically adds so little additional information that I suspect that you'll get the exact same results if you simply use an interpolated image sampled at twice the rate of the original, which is computationally far simpler than dithering. 

John

Steeve Body:
I do have drizzle 2x, 1x and no drizzle as separate stacks without bxt and with bxt. This are the first two images I posted. Is that what you mean…?

John Hayes:
Yeah, I hear what Russ is saying, but as I said, under most circumstance (like under all but the VERY best atmospheric conditions) any additional information that comes from drizzling falls somewhere very close to zero.  If it does make a significant difference, that would concern me.  That would mean that almost ANYTHING could cause noticeable swings in the output.  I'll bet that simply increasing the sampling by a factor of two might change things as much as drizzling simply based on the fact that there is virtually no difference in the information content in either case.  Remember that drizzling was developed for Hubble data where you actually can extract addition information by drizzling dithered data simply due to the fact that it's a diffraction limited system operating in a vacuum.

I'll look forward to hearing what you find...

John

Wei-Hao Wang:

Steeve Body:

Wei-Hao Wang:

Steeve Body:
I do have drizzle 2x, 1x and no drizzle as separate stacks without bxt and with bxt. This are the first two images I posted. Is that what you mean…?

No.  Not what I meant.  I meant first 2x up-sample the subs, and then do a regular stack of those up-sampled subs without any drizzling involved.  Stack and then a 2x upsample will be less effective.

I did watch Russ' video, and nothing contradicts with what I said.

Got you. Is there a module in Pixinsight that allows me to upsample all subs easily? I have always done that with photoshop…

Ha!  I am also a Photoshop person.  I think there is an "Integer Resample" function that can do 2x upsampling very fast.  But if you want to quickly do it on many images, you will need to run it on an image container.  I am not the best person to tell you how to do it, because, like you, I am a PS guy.

Steeve Body:
@John Hayes @Wei-Hao Wang  So I did another test were I resampled all the lights using integer resample in Pixinsight at 2x and re-stacked them to compare... and indeed the result are very similar like you both predicted. I still do give a slight edge to the drizzle version as it has smoother gradient and better level of contrast on the real faint detail stuff by really that is debatable at this point. An advantage of the drizzle version would be time saving in the stacking, It took considerably longer to  upsampling the lights and then stack them than just drizzle 2x in WBPP

From left to right: drizzle 2x with BXT, upsampled 2x lights with BXT, drizzle 1x with BXT

John Hayes:

Steeve Body:
@John Hayes @Wei-Hao Wang  So I did another test were I resampled all the lights using integer resample in Pixinsight at 2x and re-stacked them to compare... and indeed the result are very similar like you both predicted. I still do give a slight edge to the drizzle version as it has smoother gradient and better level of contrast on the real faint detail stuff by really that is debatable at this point. An advantage of the drizzle version would be time saving in the stacking, It took considerably longer to  upsampling the lights and then stack them than just drizzle 2x in WBPP

From left to right: drizzle 2x with BXT, upsampled 2x lights with BXT, drizzle 1x with BXT

Steeve,
I think that you might not have fully appreciated what I was trying to say—probably because I didn’t make it very clear.  Here’s what I should have said:  Drizzling beyond the seeing limit is functionally equivalent to simply interpolating the data to a higher sampling rate.  When you up-sample using IntegerResample, that’s not what you are doing.  You have to interpolate the image data to a higher sampling rate.  There are a lot of good ways to interpolate the data (using things like spline fitting) but in this case, I think that a simple linear fit will work well to make the point.  So, to do the right comparison, you need to use the Resample tool, which allows different options for interpolating the data.  As Wei-Hao said, an image container is the easiest way to do that on a batch of images.  Image containers are super simple to set up so this is a great opportunity to learn how to use them (if you haven’t already).  I’m not trying to send you back to the drawing board, but I’m pretty certain that the only reason that you see a small difference between the drizzled data and the up-sampled data is because the data wasn’t properly up-sampled.

John

Wei-Hao Wang:
Hi John,

Please see if you agree with this.  I think the best way to achieve a 2x resampling is to do it simultaneously with the image registration.  For example, one may use an image that's already 2x up-sampled as the registration reference.  Then register everything to it.  This way, each image would have only undergone one resampling.  If one first 2x up-samples and then registers, or first registers and then 2x up-samples, each image would have undergone two resamplings, which would be less ideal.

Personally I think the difference between Steeve's 2x drizzled and 2x upsampled cases would be even closer if the later is done optimally (not suggesting Steve to do it again though).  However, I expect that the drizzled case would still be better, very very slightly, perhaps so slight that only a very careful analysis can reveal the difference.  In my work flow, drizzle takes so much longer.  I no longer think it is worthwhile for the small improvement it brings (unless the original image is under-sampled, of course).

Joseph Biscoe IV:

Steeve Body:

Wei-Hao Wang:
Hi John,

Please see if you agree with this.  I think the best way to achieve a 2x resampling is to do it simultaneously with the image registration.  For example, one may use an image that's already 2x up-sampled as the registration reference.  Then register everything to it.  This way, each image would have only undergone one resampling.  If one first 2x up-samples and then registers, or first registers and then 2x up-samples, each image would have undergone two resamplings, which would be less ideal.

Personally I think the difference between Steeve's 2x drizzled and 2x upsampled cases would be even closer if the later is done optimally (not suggesting Steve to do it again though).  However, I expect that the drizzled case would still be better, very very slightly, perhaps so slight that only a very careful analysis can reveal the difference.  In my work flow, drizzle takes so much longer.  I no longer think it is worthwhile for the small improvement it brings (unless the original image is under-sampled, of course).

Are you using the WBPP script for stacking or you are doing everything manually? For me as it’s stands using the script and applying 2x drizzle works well for me without too much extra overhead on my MacBook m2 max so I’m probably going to stick to that for now.

Also would you mind clarifying that statement for me so I make sure I got this right “one may use an image that's already 2x up-sampled as the registration reference.  Then register everything to it. This way, each image would have only undergone one resampling”  

Do you mean by that pick a reference frame, upsample it 2x, register all the frames that have not been upsampled yet to it, then upsample all frames after they have been registered and stack that yeah?

This is an interesting topic. I'm curious if you got it figured out. But yes, choose a reference frame manually. Upsample it. And then use it for aligning. Aligning can be done with either the star alignment tool, or you can select a manual reference file in WBPP.