Drizzle versus upsampling

[dhalliday]

I have been trying some use of drizzle stacking...(using DSS)...
The more I read on this the more confused I get...
I am wondering if any of the Jedi could comment on what the difference is...?
I do not really understand the "process" involved in upsampling to start with...but for a long time I have been using it,and then applying a small amount of multiscaled NR.
This seems to depixilate the resulting image nicely.
On the other hand applying sharpening processes becomes trickier.
So any comment on the differences would be appreciated.
I am currently working with a lot of Ha images,at f/5...so my SNR is good.The fl in this case was 500mm...so maybe this data is not "undersampled" enough to be a valid set...
A few examples of my confusion can be seen on my Flickr site.
http://www.flickr.com/photos/daveh56/5422160996/

Dave

[Simon Hicks]

As I understand it drizzle is a process that is applied when aligned images are added together. The translation / rotation of each image relative to the other will result in one image being moved a a non-integer number of pixels (typically). So in a very simple example one image could be shifted by 0.5 pixels to the right so that the stars in both images are aligned. So you now information at more pixel sites than the original single image. You've basically made the pixel grid twice as fine in the left/right direction. You can use this information between each pixel to interpolate on a finer grid than the original single image using this 'real data'.

However, resampling just interpolates between adjacent pixels....so it does this without any more 'real data' being put into the mix.

At least I think that's what its all about.

[Juan Conejero]

Hi Dave,

Drizzle is an image integration algorithm; it has nothing to do with image resampling. Resampling involves interpolation, while drizzle is a method to combine a set of images (to be more precise, drizzle reconstructs an image from undersampled dithered data).

A good explanation of the drizzle algorithm is here:

http://www.stsci.edu/~fruchter/dither/drizzle.html

On the page above, go to the section entitled "The Method" to read a description of the algorithm. It is actually quite simple. The flux from each pixel is considered to be constrained to a smaller area (smaller than the source pixel's dimensions). That's what is called a "drop" (the blue squares on the figure on the page above). The drops from all the images being combined are averaged on a finer grid, and the resulting image has increased spatial resolution. If there are sufficient images and they are well distributed, the finer grid can be completely covered with the source data. Otherwise there will be "holes" due to lacking data. So to be really effective, drizzle requires:

- A relatively large set of images. The more the better.
- The source images must be dithered ("unaligned") at a subpixel level.

As I've said the idea behind drizzle is actually very simple, but I don't know if I'm making it easier to understand or not

As a side note, we'll have drizzle implemented as a new tool in PixInsight. Our drizzle implementation will provide the same pixel rejection algorithms as the ImageIntegration tool, and the same image registration performance as StarAlignment. Good pixel rejection is something that most drizzle implementations don't have, and this limits its applicability.

[Nigel Ball]

As a side note, we'll have drizzle implemented as a new tool in PixInsight. Our drizzle implementation will provide the same pixel rejection algorithms as the ImageIntegration tool, and the same image registration performance as StarAlignment. Good pixel rejection is something that most drizzle implementations don't have, and this limits its applicability.

Excellent news!

I look forward to this new feature 

[dhalliday]

Juan
Thanks for (always) being around.
The more I read on drizzle,...(including the article,which I have seen) the more I realized I was probably not using enough frames to get the benefit.
Also suspect I need to try it on data that is less well sampled...ie longer fl.

Getting stuff for nothing has always interested me... .....so I continue to want to use this tool.
With regards to upsampling (which I also use a lot...)...why is it/HOW is it that applying certain types of NR seems to then smooth the resulting pixelation out...?
Its a great result...but is it just cosmetic..?
And look forward to the PI version of drizzle...
Seems to me that some data should maybe have different drizzle algorithms than others.....that will be great if PI can do that.

[RobF2]

I think that helped me finally get my head around drizzle too.  Future implementation interesting too...

[zvrastil]

So to be really effective, drizzle requires:

- A relatively large set of images. The more the better.
- The source images must be dithered ("unaligned") at a subpixel level.

Hi Juan,

if I understand the article above, in order to gain something from drizzling, resolution of your optics must be at least comparable (or better) than resolution of your camera. If the limiting factor in getting better resolution is your lens (so for example your FWHM is like 3-4 pixels), drizzling can not give you much, because it only recovers information, lost by sampling to camera pixels. Am I correct?

thanks, Zbynek

[dhalliday]

Not to be too "all over the place"....but a related question;
What is the difference between "upsample"...(in Integer resample)...and "resample"....

I just played with resample for the first time... .....

Using certain parameters it seems to be doing MORE with the data.....
What determines our choice of parameters(nearest neighbor etc) in resample...?
I was playing with a 50 by 50 pixel crop; (shown here after masses of upsampling and deconvolution...)
http://www.flickr.com/photos/daveh56/5425722433/

Sorry to be pestering...

Dave
PS going back and looking a bit more at this "resample" function.....maybe my question is just too complicated for a simple explanation....
So no problem if no one volunteers one...
But it seems almost as though this is applying upsampling AND some NR.....

[dhalliday]

With regards to "Resample"...just saw this...
http://pixinsight.com/forum/index.php?topic=556.0

It works well but I need to update my PI version...
I find the ACDNR does not work so well on binned data...Maybe it is just too noisy somehow.
Maybe greycstoration is better...

The sky is easily mottled...
Resample seems better
http://www.flickr.com/photos/daveh56/5467182640/#/photos/daveh56/5467182640/lightbox/

Any comments  appreciated.

[NKV]

As a side note, we'll have drizzle implemented as a new tool in PixInsight. Our drizzle implementation will provide the same pixel rejection algorithms as the ImageIntegration tool, and the same image registration performance as StarAlignment.

Juan, any news? One year passed...

[Carlos Milovic]

Hi
News from here... I'll implement a different approach to this, called Super-resolution. In a nutshell, it is a combination of registration, integration, deconvolution and noise reduction.
Right now I'm working on a new automatic registration method, that can be applicable to any kind of images, but that only calculates rotations and shifts. A local registration algorithm may be used on top of this. The benefit from this rigid technique, is that it should be fast and more accurate than other methods (for example, fourier analysis). Also, this should be enough for long exposure frames, where distortions frames are irrelevant. A more local approach should be used for solar, lunar or planetary images, that may work over the rigid solution.
After the registration routine, I must write the deconvolution/noise reduction steps, that use all the data from the series, and yields an optimised solution.
I'll post the modules somewhere as soon as I have something to try with real data. I'll be quite overloaded with work at the university the following weeks, so I may release a first version in July.