Deconvolution in PixInsight - a written workflow

[steppenwolf]

I have been working to get a repeatable workflow going for deconvolution of my 3nm narrowband images using PixInsight and having achieved that, I decided to document my process for future reference. I thought it might be useful for other PixInsight users just starting out so here it is. As with all things PixInsight, do try experimenting with different settings, especially if you want to deconvolve wideband Luminance data.

Deconvolution Extract.pdf

[PhotoGav]

Excellent job Steve, thank you for sharing. Your tutorial session at SGL SP2018 (the Wet One) on exactly this method is the closest I have come to being pursuaded to buy PixInsight. One day....

[RayD]

Super job.  Thanks Steve, that's a really useful document  Thank you for taking the time to produce it.

Edited November 20, 2018 by RayD

[ollypenrice]

I'll be studying this with interest, Steve, since I have never been able to get any sense out of deconvolution (other than with a small format 8 bit system in an old copy of Astra Image which I once had.)

Thanks for posting.

Olly

Edit Corrected the name of Astra Image.

Edited November 21, 2018 by ollypenrice

[tony210]

Thanks Steve -very useful - my Decon in Pixinsight has given very  variable results-Tony.

[wimvb]

That's a great write up, Steve. Thanks for sharing this.

Just a few minor remarks.

1. While it is possible to apply deconvolution without a luminance mask (you have to play around a lot with the Regularisation parameters and dark ringing parameter), imo, it usually pays off to use a Luminance mask with strong protection of the background. Just applying the STF as a permanent stretch, doesn't always give enough protection.

2. If you want to avoid ringing alltogether, make sure you have a preview with smaller stars over bright nebula. This is usually the first area in an image to show ringing. And if the plan is to use HDR processing later on, that will enhance any ringing artefacts from deconvolution.

[steppenwolf]

8 hours ago, wimvb said:

1. While it is possible to apply deconvolution without a luminance mask (you have to play around a lot with the Regularisation parameters and dark ringing parameter), imo, it usually pays off to use a Luminance mask with strong protection of the background. Just applying the STF as a permanent stretch, doesn't always give enough protection.

Thanks Wim, I think I covered that in item 8 of the Luminance Mask but as you say, it is an important step because deconvolution can pick up on background 'noise' and produce additional artefacts.

8 hours ago, wimvb said:

2. If you want to avoid ringing alltogether, make sure you have a preview with smaller stars over bright nebula. This is usually the first area in an image to show ringing. And if the plan is to use HDR processing later on, that will enhance any ringing artefacts from deconvolution.

Yes, it is important to have several previews to test so that you can check the interaction between the different components of the image. The biggest problem that I have had in the past is with the brighter stars but the star mask has resolved this. However, I must work on a way of reducing the size of larger stars as well as they still bug me!

[sloz1664]

Thank you Steve, a most informative read and one I will apply to my future Pixinsight processing procedures.

Steve

[wimvb]

2 hours ago, steppenwolf said:

However, I must work on a way of reducing the size of larger stars as well as they still bug me!

If the stars are overexposed, and have a "flat top", while still in the linear stage, deconvolution can at best, only decrease the halo somewhat. (I have an idea about a possible method, but must test it first. It involves going against "best practices" for deconvolution.) Morphology transformation may be a better tool for reducing large stars.

If the number of regularisation (now, there's a word for scrabble) layers is increased, and their levels, it's possible to target only stars in deconvolution. Also it helps if the number of iterations is increased and dark ringing protection kept to a minimum. This needs to be done with a star mask that only targets those large stars.

As you wrote, experimentation is required.

 

[steppenwolf]

29 minutes ago, wimvb said:

If the stars are overexposed, and have a "flat top", while still in the linear stage, deconvolution can at best, only decrease the halo somewhat. (I have an idea about a possible method, but must test it first. It involves going against "best practices" for deconvolution.)

I'd be really interested to hear how you get on with that, Wim.

[souls33k3r]

Awesome stuff Steve.

This is why i love this hobby, you find gems who are happy to share their knowledge and experience only to see others succeed.

Top marks to you sir

P.S: Mind if i ask, what book is this from? I ask this because it's got page numbers on it.

Edited November 21, 2018 by souls33k3r

[steppenwolf]

4 minutes ago, souls33k3r said:

P.S: Mind if i ask, what book is this from? I ask this because it's got page numbers on it.

This is a draft extract from a 'book' that I am writing purely for my own reference at the moment and currently have no plan to publish!

[souls33k3r]

4 minutes ago, steppenwolf said:

This is a draft extract from a 'book' that I am writing purely for my own reference at the moment and currently have no plan to publish!

Ah ok. An excellent write up. 

[RayD]

20 minutes ago, steppenwolf said:

currently

Aha, I shall hang on to this word in hope ?

[souls33k3r]

1 minute ago, RayD said:

Aha, I shall hang on to this word in hope ?

The process workflow is for PI not PS ?

[steppenwolf]

9 minutes ago, souls33k3r said:

The process workflow is for PI not PS ?

Yes indeed it is. I do all my preliminary processing - calibration (bias, darks and flats), registration (alignment), integration (stacking) and deconvolution - in PixInsight and then transfer the data across to my old friend PhotoShop where I do everything else.

[souls33k3r]

I seriously need to learn PS 

[steppenwolf]

1 minute ago, souls33k3r said:

I seriously need to learn PS 

There are several schools of thought on all of this - of course! PixInsight will process an image from start to finish and it is aimed specifically at astophotography so this has some appeal but this appeal diminishes somewhat if you already have experience of any kind of image processing using PhotoShop! PhotoShop is very intuitive and everything happens in proper real-time so you can very easily see what is going on and make adjustments accordingly, whereas PixInsight is much more mathematically based and, frankly unfathomable to me at times! However, there is more than enough room for both and currently, while I slowly get to grips with PI, I use the best attributes of each.

[souls33k3r]

2 minutes ago, steppenwolf said:

There are several schools of thought on all of this - of course! PixInsight will process an image from start to finish and it is aimed specifically at astophotography so this has some appeal but this appeal diminishes somewhat if you already have experience of any kind of image processing using PhotoShop! PhotoShop is very intuitive and everything happens in proper real-time so you can very easily see what is going on and make adjustments accordingly, whereas PixInsight is much more mathematically based and, frankly unfathomable to me at times! However, there is more than enough room for both and currently, while I slowly get to grips with PI, I use the best attributes of each.

You've made some really good points about both and i will agree to that but what you said is exactly the reason what makes me want to learn PS even more which is PS is very intuitive and everything happens in real time. PI does have the preview option but most of the times you work on small previews which can be a little time consuming but the whole mathematics piece most of the times goes above my head. 

[steppenwolf]

3 minutes ago, souls33k3r said:

PI does have the preview option but most of the times you work on small previews

Yes, in my earlier days of exploration, we used to call this a 'kludge' That said, as you can see from my article, I use the previews to good effect when testing deconvolution values but for stretching and image?

[souls33k3r]

2 minutes ago, steppenwolf said:

Yes, in my earlier days of exploration, we used to call this a 'kludge' That said, as you can see from my article, I use the previews to good effect when testing deconvolution values but for stretching and image?

Going through it right now. 

I'm very picky about my read and that's because i do not want to lose interest by being bored with jargons and whatnot but you sir have a knack for writing. #Captivated.

I'd happily proof read your "drafts"

[wimvb]

1 hour ago, steppenwolf said:

This is a draft extract from a 'book' that I am writing purely for my own reference at the moment and currently have no plan to publish!

Just keep those draft extracts coming.

[Laurin Dave]

Thanks Steve for your excellent write up,  I shall add it to my copy of your Dark Art or Magic Bullet book which arrived yesterday  .. I did notice that in this post Deconvolution now starts on page 51 rather than 45 as in the book.. What other gems are you working on?

Dave

[steppenwolf]

6 minutes ago, Laurin Dave said:

Deconvolution now starts on page 51 rather than 45 as in the book.. What other gems are you working on?

There is no significance to the page number as this 'chapter' isn't in any way associated with either 'Dark Art or Magic Bullet?' or 'Making Every Photon Count' - this is a chapter from my own personal reference manual (I do like to document the stuff I learn as it makes it easier to absorb and I can easily refer to it in the future!) and isn't currently being considered for publication. ?

I hope you enjoy 'Dark Art'.

[wimvb]

11 hours ago, steppenwolf said:

I'd be really interested to hear how you get on with that, Wim.

Somewhat less in a rush, I've had time to think about this. My initial thought was to use a gaussian PSF for deconvolution (because the largest, brightest stars have more of a gaussian shape). But when I thought about it I realised that bloated stars have a plateau, where deconvolution will cause a ridge to form, as in the attached image. On the left is the bloated star; on the right is the deconvolved bloated star. The intensity of the deconvolved star is 0.99 on the ridge, and 0.96 in the middle. Btw, this image isn't stretched, it shows the star in its linear state. But I did resample the image 400 %.

The only way to avoid this ridge is to make sure that the star is less bloated. This can be done by slightly blurring it. With AtrousWavelets or MLT, remove the first two layers (using a star mask of course) This doesn't make the star smaller. But it will give it a slightly rounder top, which will tolerate deconvolution a little better. Since all this is done in the linear stage, differences will be very small. When I tried this, I had to experiment with the regularisation settings (increasing layers to 3, and increasing the noise threshold for all layers). I didn't particularly like the resulting star shape, even if it was visually a little narrower.

in the end, this method for keeping stars under control is probably easier:

http://www.pixinsight.com/tutorials/NGC7023-HDR/index.html#High_Contrast_Small_Scale_Structures

 

Edited November 21, 2018 by wimvb