Startools Deconvolution

[Startrek]

I read a comment on another forum in regard to Deconvolution and wanted to know if Startools Deconvolution has address the technical aspect of the comment as follows -

Comment......
When an image is stretched you cannot do a proper Deconvolution on it.You maybe able to improve the stretched image with something that resembles Deconvolution but calling it Deconvolution doesn’t make it so.Proper Deconvolution relies on an accurate measure of the PSF and a stretched image loses the property of a unique PSF.
You do a Deconvolution on a stretched image if you want but to improve the image but mathematically it won’t be a proper Deconvolution.

I welcome any comments in regard to the above
PS: If Ivo is around I would appreciate his expertise on the subject as well
Thanks

Martin

[Stefan B]

Hi Martin,

my understanding is that SVD deconvolves the linear data, but you can see the effect on the stretched data thanks to the tracking feature. Although the documentation reads a bit different:

It preferably operates on highly processed and stretched data (provided StarTools' signal evolution Tracking is engaged)

So I might be wrong on that one.

Regards
Stefan

[Mike in Rancho]

Hi Martin,

Was the comment in regards to ST, or just general discussion?

In any event, ST puts the math in the right place. It's all about the tracking and time-shifting, and even the cross-talk between modules looking fore and aft. Really quite cool.

I'd have to go review the docs again, but Ivo has also separately explained how the deconvolution is done on the linear data, a feature that goes way back I'd say. There's also a good internet video interview where time shifting is explained rather well.

And fairly recently here there was more discussion with Ivo on the time-shifting and tracking, where we learned about the numbers and math behind the bias sliders in color, which is also performed on the linear data (and thereby the numbers make sense with what is being seen). It may have been the same thread where Ivo explained how SVD operates on the linear state, due to tracking, even though through time-shifting you are seeing the live results of those calculations, and the forward and backward-looking also allows incorporation of other modules you have applied.

Unless I messed up my own understanding of it all, of course.

[Startrek]

Thanks Stefan and Mike,
The comment was from a person who has never reviewed or used Startools
I’m sure your both correct in regard to Decon
A comment from Ivo would be icing on the cake
Thanks again
Martin

[admin]

What Mike said.

StarTools' deconvolution implementation is one of the major examples of Tracking, and why you are able to achieve so much better results in ST than in any other software currently available. This page on the website explicitly mentions the exmple of how deconvolution is augmented by signal evolution Tracking.

What others often confuse for "guardrails" in ST, is really a very complex signal flow piece, that was specifically designed to achieve superior signal fidelity (and by extent - detail and control).

[Startrek]

Ivo,
Thanks
Excellent link and sums everything up , especially Decon

Cheers
Martin

[Mike in Rancho]

Thanks Ivo, I forgot about that page, which does explain things quite well.

But it does get me thinking about issues with using ST's handful of re-do/restore options -- are the forward and backward looking time travel tie-ins and adjustments unraveled, or further adjusted for, when those are employed?

For example, the page states that SVD knows how you are going to stretch. But what if you hit keep in SVD, being happy with your results there, but then maybe after Color you go back to AutoDev re-do global stretch for a little adjustment?

Likewise with Restore, where there are option to retain your deconvolution, but was that deconvolution taking into account things that have no longer exist?

Not that I've been able to notice any flaws, but just curious as to usage, and knowing when it might be best off to just start from scratch, or use Restore just to original or original + wiped.

[admin]

Apologies - I missed this (lots going on over the past few weeks)

But it does get me thinking about issues with using ST's handful of re-do/restore options -- are the forward and backward looking time travel tie-ins and adjustments unraveled, or further adjusted for, when those are employed?

For example, the page states that SVD knows how you are going to stretch. But what if you hit keep in SVD, being happy with your results there, but then maybe after Color you go back to AutoDev re-do global stretch for a little adjustment?

The deconvolution will indeed have been based on the previous stretch. This is largely the reason why there is a recommended workflow, rather than an "anything goes" advice. That said, unless extremely/dramatically different, a global re-stretch specifically, should not impact the fidelity of the result much. That is mostly because it is actually the local changes in finer detail (as caused by Contrast, HDR, Sharp - all of which are preserved when you redo the stretch specifically) that are the major contributors to noise/artifact exacerbation.

Likewise with Restore, where there are option to retain your deconvolution, but was that deconvolution taking into account things that have no longer exist?

Indeed.

However, the effect on the end result - as you noticed - may still not be dramatic, largely because throughout your subsequent "new" workflow, any new changes that may build on the deconvolved result, are ultimately measured (and subsequently throttled) against the linear data as well. It just means that other modules (in so far they use Tracking) will now "compensate" for decon having been, perhaps, a bit overaggressive in some areas for the new stretch.

Not that I've been able to notice any flaws, but just curious as to usage, and knowing when it might be best off to just start from scratch, or use Restore just to original or original + wiped.

original + wiped is indeed your best bet to be "thorough".

[Mike in Rancho]

Thanks Ivo.

Yes sometimes the differences do seem to be quite subtle, if you can even see them, but I do like trying to help ST do what it does best.

As a semi-tangential matter, is SVD using its "peek ahead/peek behind" skills just for noise and artifact tracking, or also to derive clues for the spatial PSF modeling? I am unsure if I explained it wrong in our current CN challenge side-thread - after Eyal said all software uses a single PSF for the entire image.

[admin]

As a semi-tangential matter, is SVD using its "peek ahead/peek behind" skills just for noise and artifact tracking, or also to derive clues for the spatial PSF modeling?

It does, which is actually the limiting factor in how much deformation can be fixed. ST has to balance noise (and thus artifacts) propagation due to noise in the signal, and that signal "moving" to the location it should be sitting. The "moving" happens over multiple iterations.

It's what the "Spatial Error" parameter is for; it allows you to say "I don't care about the precise precision of the location" (it's allowed to be a bit fuzzy), as long as the signal "moves" more.