Deconvolution in PixInsight - a written workflow

[steppenwolf]

Thank you for the link, Wim, I will have to study that one very carefully!

[Big Jim Slade]

I must admit I'm a bit confused about the PSF part here. I don't quite understand why only using just Moffat (in Steve's example) or just Gaussian as Wim mentions above, is necessary. To get the most accurate fit to characterise the PSF don't you need both?

(I should add of course: thanks to Steve for doing this. I've been through multiple write-ups on Deconvolution and it's very hard to find one that makes a very complex process as clear as this does.  )

 

Edited November 21, 2018 by Big Jim Slade

[wimvb]

6 hours ago, Big Jim Slade said:

I must admit I'm a bit confused about the PSF part here. I don't quite understand why only using just Moffat (in Steve's example) or just Gaussian as Wim mentions above, is necessary. To get the most accurate fit to characterise the PSF don't you need both?

(I should add of course: thanks to Steve for doing this. I've been through multiple write-ups on Deconvolution and it's very hard to find one that makes a very complex process as clear as this does.  )

 

The psf process models stars with a mathematical function. It assumes that optics, atmosphere, seeing, and tracking together, produce a bell shaped star profile. For small and medium sized stars, the exact function is a little pointier than a gaussian function (traditional bell shape). But when the camera starts to saturate, (and maybe with a high level of oversampling?) the profile becomes less pointy, and resembles a gaussian function. Deconvolution tries to reverse the effect of seeing, not camera saturation. That's why moffat is generally better. Otoh, you can always try to use a tool outside it's intended scope. That's why I thought that it might be used to correct stars that become bloated. But it just can't be used on overexposed stars.

[steppenwolf]

4 hours ago, wimvb said:

The psf process models stars with a mathematical function. It assumes that optics, atmosphere, seeing, and tracking together, produce a bell shaped star profile. For small and medium sized stars, the exact function is a little pointier than a gaussian function (traditional bell shape). But when the camera starts to saturate, (and maybe with a high level of oversampling?) the profile becomes less pointy, and resembles a gaussian function. Deconvolution tries to reverse the effect of seeing, not camera saturation.

This would be my understanding of why I only use Moffat sample as well.

[RichLD]

Excellent write up Steve, many thanks!

[Scott]

A very confusing process made possible thanks to your ability to communicate via the written word. I shall give it a go when I finally get some data worthy of more than just a quick levels/curves tweak.  

Thanks for taking the time to experiment with this and then sharing your findings.

[steppenwolf]

16 hours ago, RichLD said:

Excellent write up Steve, many thanks!

My pleasure!

15 hours ago, Scott said:

A very confusing process made possible thanks to your ability to communicate via the written word.

I hope you find it useful, Scott - it's working every time for me now whereas before, I found deconvolution a bit of a variable process!

[alan4908]

An excellent write up Steve. 

The only major difference from my own workflow relates to the star selection process for the PSF creation. So for me:

1.  I constrain these to be within my camera's linear region, in addition to having a minimum signal level,  so for me, the max value is about 0.6. 
2.  I never (try) to pick stars that are embedded within nebulosity since their stellar profile will most likely be contaminated by the nebulosity.  
3.  To minimize optical distortions I try to pick stars that are near the center of the frame.

After you have created a list of candidates, Adam Block's Pixinsight deconvolution tutorial advises that you should sort the Mean Absolute Deviation (MAD) column and look at the numbers.  This gives you an idea of the quality of the input data.  Apparently, what you are after is a consistent set of numbers within a factor of say x2.  If you find something outside this range then you should consider deletion of the outliers.

Alan

[steppenwolf]

18 minutes ago, alan4908 said:

After you have created a list of candidates, Adam Block's Pixinsight deconvolution tutorial advises that you should sort the Mean Absolute Deviation (MAD) column and look at the numbers.  This gives you an idea of the quality of the input data.

Thank you for that, Alan, I shall have a look on t'Internet for that tutorial!

[wimvb]

2 hours ago, alan4908 said:

The only major difference from my own workflow relates to the star selection process for the PSF creation. So for me:

1.  I constrain these to be within my camera's linear region, in addition to having a minimum signal level,  so for me, the max value is about 0.6. 
2.  I never (try) to pick stars that are embedded within nebulosity since their stellar profile will most likely be contaminated by the nebulosity.  
3.  To minimize optical distortions I try to pick stars that are near the center of the frame.

These are sound selection criteria.

1. I try to avoid stars with obviously low amplitude, and always avoid stars with an amplitude larger than 1 (note that the amplitude value is a calculated one; real values are always < 1 ). But the latter usually have gaussian profile anyway. During the selection part of psf, I go for stars that just are barely visible in the linear image, with stf turned off, and those slightly dimmer.

2. I haven't found this to be a real problem, and sometimes it can be difficult to avoid those stars. But if you can create a reliable star profile without them, the better.

3. Always good practice. The optics will determine which part of the image works best. You'd want to avoid any stars with obvious eccentricity issues.

[OhNo]

I'm late to the party, but want to thank steppenwolf for posting this gem of his.

Being a new PI user does make the art of processing a pretty vertical learning curve! LOL At least when I go back and use other tools like StarTools or PS they seem kind of easy to use!

[Datalord]

Since this thread was necroed and I now saw it, I want to mention that we now have an amazing starmask tool at our disposal in the form of Starnet++ for PI. This process, set to a stride <=64 produces starmasks that are way, way better than anything I can get out of the StarMask process. On the resulting starmask, you can do a morphological transform with a dillude, followed by a MultiscaleLinearTransform with bias disable for scale 1 and 2 and you have the prettiest starmask with soft gradients you can imagine.