Mastering (?) Deconvolution

[dhalliday]

Group
I wondered if a little "chat" could be got going on this tool...
I am a big fan of M1,M57 etc..
Small stuff at 1300mm..
I often use "image upsample",drizzle (in DSS) to "cheat" a bit
I have some success with HDR wavelets...(but it KILLS the color..)

But I think Decon is "the key to the kingdom..

I was hoping we could all chip in and discuss/share around this tool.
My FIRST lesson from Juan was that this should be used BEFORE I stretch my data...
Is this correct?
So if I have "screen transfer" ON...(to see what the Decon is doing to a preview)...this is the way to go?


Second question...
It only has 38 parts...(!!)
What the heck do all those sliders do..?
I mean is there a SYSTEMATIC way to approach each one?
Third question...
Is this best done on a RBG or just a Lum image...and is it better to apply a mask to the low SNR areas while using it.??
And how should we create a Lum mask on a (virtually) black image...

Hoping for a discussion,or advice..
Here is some post deconvolution results on M1 (Lum dat-I confess !! )
just to give an idea what I am working with...
http://www.flickr.com/photos/daveh56/4294849972/sizes/l/

I think getting MORE photons into the image is a good start,but then what...?

Dave

[vicent_peris]

I have some success with HDR wavelets...(but it KILLS the color..)

Just a little note here. HDRW doesn't kill color. It substracts the large scale color structures of the image; this is why the image seems to have a muted color. But, if you raise the color saturation, you will see that, in fact, color has been boosted for the little structures.


Vicent.

[Harry page]

Hi

yep i would like to help but i am useless at this as well, but a couple of things I do know

1) decon only works well on high SNR areas , so i do not think your image is suitable for decon ( I think  )

2) Following Item 1 all areas of low signal ( ie background ) would benefit from masking


Thats it , my lack of knowledge all laid bare   


harry

[dhalliday]

OK guys...that is a start...
I will return to check my traps in 24 hours...

[dhalliday]

Hmmm
Must be using the wrong bait...

Well my first question is just to RECONFIRM what Juan said;
Deconvolution should only be done on "linear" data...
That means unstretched,as it comes out of DSS (or whatever..)
Correct ?

This just seems odd...because the way I have been using it...on data that has had a stetch,DBE (now,WAS ABE !) and NR...
It has seemed to help...depending on those two darn sliders...but it certainlt adds noise to the image...can do quite a number on my stars...
I am thinking maybe it should be done with a Lum mask ?
Again that seems hard to set up on early/unproccessed data
The "early" data looks pretty fuzzy..and NOISY...(I guess that's the same thing...)
So comments here are what I am looking for,for starters.

Dave

[Niall Saunders]

Hi Dave, OK, this will be a good challenge for my memory, AND my 'understanding' of Deconvolution.

First - think of the 'ideal' world - a star is a 'pinpoint' light source and, as such, it can ONLY illuminate ONE pixel on your CCD sensor.

A matrix representation of this might be
0 0 0 0 0
0 0 0 0 0
0 0 1 0 0
0 0 0 0 0
0 0 0 0 0

However, in the 'real' (theoretical, still) world, we have diffraction effects of the optics, so the 'best' you can then get is the size of the 'Airy Disk'. In other words, your 'pinpoint' has been 'convolved', or 'spread out'. Now the matrix representation might be
0 0 0 0 0
0 1 1 1 0
0 1 3 1 0
0 1 1 1 0
0 0 0 0 0

When you then add in atmospheric distortion, you might then see
0 1 1 1 0
1 2 3 2 1  
1 3 5 3 1
1 2 3 2 1
0 1 1 1 0

The aim of deconvolution is to figure out a way of creating an 'inverse matrix' such that what you 'actually see', can be made 'more like' what you 'actually want'.

However, if you have 'stretched', or otherwise modified, your original raw data in a 'non-linear' way, then the deconvolution algorithm does not really have a chance of creating the inverse matrix, because it CANNOT know what YOUR 'extra' transformations were. The HUGE advantage of PixInsight is that you can perform a whole host of LINEAR transformations to your data, and still hope to be able to figure out a deconvolution 'kernel' (or matrix).

OK, if you weren't going to have much success in establishing a deconvolution kernel without those transformations, then those linear transformations wouldn't make things any 'worse'. And, of course, the MAIN non-linear transformation, the Histogram Curve that actually allows you to 'see' what is in your image, doesn't have to be applied until much later on in your workflow, because you can initially use the STF 'visualisation' transfer to 'see' the image, without actually 'changing' the LINEAR data needed by Deconvolution.

Clear as mud  

Cheers,

[sleshin]

Actually, Niall, a good explanation. Thanks for posting it.

Steve

[dhalliday]

Niall
Thanks!...so I will give it a go under STF..
So I guess I cannot have already applied NR either...?
What about ALL those sliders/settings...whats THAT all about

Dave

[Niall Saunders]

Hi Dave,

Nope - NR affects the image in a 'non-deconvolution-friendly' way.

I don't pretend to fully understand the principle of deconvolution - well, I do, but not yet well enough to explain it in intricate detail. However, if you are interested, the best treatise I have come across on the subject was in The Handbook of Astronomical Image Processing (HAIP, the book that accompanies the AIP4WIN software package). I really cannot emphasise enough how useful this book is. A heck of a lot of the drivel that I spout here (and elsewhere on the 'net) has come as a result of what I learbed reading that particualr 'bible'.

And, as for the sliders. the set the amount, colour, and flavour of the 'magikery' that Juan uses to make his processes work. Quite simple really. Keep twiddling them until your image looks right, then email it to the guys on the APOD committee 

Honestly, I have NEVER managed to get deconvolution to work, but I do now realise that I more than likely NEVER actually have 'good enough' data in the first place (deconvolution will NOT work with 'poor' data - it just doesn't stand a chance). But, one day, I will get a good night, perhaps with a higher resolution camera, perhaps just with the DSI-II, and the data will 'look good' right after the basic stages - and then I might throw some serious effort into making deconvolution work.

Until then, I am waiting for Harry to 'show me how' !!

Cheers,

[Harry page]

Hi Niall

I hope you are a very patient man , as you are going to waiting a very long time 


Harry

[Niall Saunders]

No hHarry, that is NOT how it works.

As Video-Jedi (Vedi, if you like), you get the statutory 90 minutes to produce the video, as requested. Sorry, but your's is a position of responsibility.

We are all looking up to you now (and I am deleting my two other tutorials as I type)

Cheers,

[Harry page]

i am sorry I will not be reading the last email for at least a year 


Harry

[Niall Saunders]

84 minutes

Less typing, more videoing . . . . .

[dhalliday]

Does Juan have a tutorial on it ? I have not seen one...
I am not much further ahead...
Niall;how does one asses the "quality" of ones data..??...(thinness of the histogram/noise level..?)
Is there a cursor readout one can scroll over the background to asses "noise"..?
?does Decon (and HDR for that matter) need to be applied in "passes"...
I mean one at layer 3,and another at layer 4 etc etc...just wondering...

How do we induce Juan to chime in here...if we are stuck??

Dave

[Niall Saunders]

Hi Dave,

"assess quality of data" ? I don't really have the answer for that - my DSI-II data has just never been 'good enough', that's all I can say.

So, for now, I just ignore Decon - I get an image that I am happy enough with when I don't bother with Decon.

Juan has discussed Decon here on the Forum (if memory serves me right), but I am on my iPhone just now, so searching is just too much hassle

Have a hunt around for yourself.

Cheers,