How do you apply noise reduction and deconvolution masking?

[RobF2]

How do people usually do their noise reduction and deconv?
Specifically, do you just use the masking functionality within these processes, or do you go for the extra control of Starmasks?

I struggle a lot to get optimal settings, and have only just realised I should probably be doing global masking with Starmasks before I fire up the individual processes.  

I imagine in the case of deconv if you have masked much of the low signal/noise part of the image you could be more aggressive and run a lot more iterations along with separate passes (e.g. settings designed to bring out large features, followed by additional pass and settings for small features (reliant on quality of data too of course)).

Would appreciate hearing from the gurus.....

[Nocturnal]

Hi,

I generally do noise reduction with ACDNR and its built in luminance mask. Generally you want to apply NR only in low SNR areas and those are generally dark so this works quite well. I see no reason to build a star mask as this would apply NR in your targets and you don't necessarily want that.

- histogram
- ACNDR, luminance masked, generally small stddev if the image allows
- another histogram to clean up the empty tail in the histogram created by ACDNR

I never do deconvolution so I can't help you there, sorry.

[RobF2]

Hi,

I generally do noise reduction with ACDNR and its built in luminance mask. Generally you want to apply NR only in low SNR areas and those are generally dark so this works quite well. I see no reason to build a star mask as this would apply NR in your targets and you don't necessarily want that.

- histogram
- ACNDR, luminance masked, generally small stddev if the image allows
- another histogram to clean up the empty tail in the histogram created by ACDNR

I never do deconvolution so I can't help you there, sorry.

Thanks Sander - just wondering what you mean by "another histogram to clean up the empty tail in the histogram created by ACDNR"?

It also seems tantalising the extra info that deconv can tease out, so interested to hear from others too....

[Nocturnal]

Rob,

simply try the steps I outlined and you'll see what I mean when you do the second histogram. My description is probably poor. The NR reduces the number of dark values being used so you can clip the backpoint quite a bit while only throwing away a few pixels.

[dhalliday]

I am a Decon junkie..
But it needs a decent SNR to stay clean.
You need to dering,and mask the background...
The "blur" function can rescue bad stars,,,to a degree.
It takes a TON (?10 tons) of experiment/practice...

I have applied up to 50 iterations on stuff like M57...and maybe only 3-5 on weaker stuff.
If it imroves the detail...use it.
Often,like unsharp...it just makes noise..
I applied only a whiff here;
http://www.flickr.com/photos/daveh56/4836256881/

And a ton here;
http://www.flickr.com/photos/daveh56/4683696025/in/set-72157620881341430/

Ideally you would mask the stars AND background...
Then work on the stars seperately under a star mask
The mask is just the (extracted) lumunance,which I stretch and clip a bit.
In ACDNR you have the mask feature built right in..
In decon you have to make it from scratch...

[RobF2]

Rob,

simply try the steps I outlined and you'll see what I mean when you do the second histogram. My description is probably poor. The NR reduces the number of dark values being used so you can clip the backpoint quite a bit while only throwing away a few pixels.

Thanks Sander - will check it out....

[RobF2]

I am a Decon junkie..
But it needs a decent SNR to stay clean.
You need to dering,and mask the background...
The "blur" function can rescue bad stars,,,to a degree.
It takes a TON (?10 tons) of experiment/practice...

I have applied up to 50 iterations on stuff like M57...and maybe only 3-5 on weaker stuff.
If it imroves the detail...use it.
Often,like unsharp...it just makes noise..
I applied only a whiff here;
http://www.flickr.com/photos/daveh56/4836256881/

And a ton here;
http://www.flickr.com/photos/daveh56/4683696025/in/set-72157620881341430/

Ideally you would mask the stars AND background...
Then work on the stars seperately under a star mask
The mask is just the (extracted) lumunance,which I stretch and clip a bit.
In ACDNR you have the mask feature built right in..
In decon you have to make it from scratch...

Thanks David.  I see your equipment is somewhat similar to mine in some ways, so that's very handy to hear.  I'm sort of kicking myself I didn't tweak to using masks with deconv a bit sooner.  Your post also makes me realise I've been judging deconv efficacy by what it does to the stars first, and faint fuzzies 2nd - which isn't always going to be the ideal solution of course. 

The whole concept of deconvolution fascinates me.  I always feel a bit like the TV depiction of an FBI agent rescuing the numberplate of the suspect from a dodgey camera shot  8)

[Niall Saunders]

Hi Rob,

I always feel a bit like the TV depiction of an FBI agent rescuing the numberplate of the suspect from a dodgey camera shot

Isn't that exactly what we all are? Federated Brethren of Image-processors ?

Just keep in mind what 'Deconvolution' is trying to do for you. What I have always understood the process to involve is that we are saying,

"We know what our original image (of a point source, like a star for example) 'should have looked like', and we can postulate a series of understandable and definable steps that have corrupted that 'ideal' to give us the (star) image we are looking at now. So, all we need to do is work out the 'inverse transformation' and apply it to our 'actual data' in order to recover the 'original data' from it. And if we know what this inverse transformation is for a star, then it should be exactly the same for the remainder of the image"

(with sincere apologies to the HAIP for extreme over-simplification )

Of course, the challenge then becomes defining the inverse transformation, and hoping that this transformation is apllicable not just for one star in one area of the image, but for all objects, in all areas of the image. This is an INCREDIBLY complicated process - just try working with a simulated image, even with only a few stars, which has then had all the usual 'noise sources' thrown on top of it. You just CANNOT get the original data back - you might get 'some', or even 'a lot' - but you will have a whole bunch of 'hangers on' that come along for the ride (just like weeds in the garden, you get rid of one lot, turn around, and something else has appeared behind you!)

And, you have to remember that HOW you set up your Deconvolution process is almost certainly going to be completely different for EVERY image that you work with. Like all tools, imaging or otherwise, perfection usually only happens in the hands of a skilled artisan.

Going back to your TV and FBI example - don't you now enjoy a wry smile to yourself whenever you see Jack Bauer or Chloe zoom in to read a credit card number from an orbiting satellite hundreds of miles away, above a layer of city smog - and do this IN REAL TIME

Bring on the PixelPolice    
Cheers,

[RobF2]

Yikes - I always wanted to belong to Bureau Nialls - cool!

Yes, deconv is certainly an adventure everytime you use it, but I love the bit of extra crispness it gives Lum or Ha stacks providing there is plenty of good data to start with.  I understand why you need to tweak and get optimal parameters for each image, but I do seem to be using similar values a lot of the time with my rig.  I might even have a go at saving icons for "starting parameters" for fine and large scale detail - still getting my head around the size and shape functions for the PSF, but its great to have so much control.