RC BlurXterminator …. Any comments ?

[Mike in Rancho]

Cool. I gave it a quick test on my poor, low integration 1-night data of M81/82, since I already have so much stuff in the background annotated. Even cranked it up to 20 iterations, centroid tracking, and turned off deringing momentarily.

A pretty artifacted mess, as you can imagine. I was binned at 50% and suppose I could have left it native.

But indeed, little blurs, even tiny ones, retained their blobby distant galaxy shape. Only one I saw, a PGC that started out star shaped to begin with, got a bit brighter in the center, but still not exactly like a nearby actual star of about the same beginning brightness.

The only places I see of possible breakdown are the cores of larger galaxies, here M81, which seemingly can be star-ized. Actually pretty common, and I can't think so many such galaxies just so happen to have a foreground MW star dead-center.

[admin]

The only places I see of possible breakdown are the cores of larger galaxies, here M81, which seemingly can be star-ized. Actually pretty common, and I can't think so many such galaxies just so happen to have a foreground MW star dead-center.

That would be expected behaviour and be essentially a correct representation. Spiral galaxies tend to have very bright cores (it's a very violent place with lots of stars incredibly densely packed together). They do not tend to be diffuse (lots of exceptions to that rule of course as always ). These galaxy cores will very quickly over-expose;

(random linear luminance M81/M82 dataset at 360s exposures)

In cases where the cores do not over exposure, however then - as opposed to BXT - the "star-ization" of galaxy cores in this way is completely controllable in terms of dynamic range. If they did not over-expose in the linear dataset, then they don't have to over-expose in the SVDecon image. That's because true decon (as opposed to BXT apparently) increases dynamic range;

BXT appears to treat stars as morphology ("shape") and size problems. But true deconvolution actually makes stars brighter (it re-concentrates energy). In ST, you can use Dynamic Range Extension parameter to give features that would otherwise over-expose the dynamic range "breathing room". E.g terrible dataset of M8;

Without decon;

Selection_785.jpg (87.98 KiB) Viewed 33476 times

With decon (with what I assume is what you call "star-ization" of the core)

Selection_786.jpg (87.53 KiB) Viewed 33476 times

With decon + 20% (1.20) dynamic range increase

Selection_784.jpg (82.75 KiB) Viewed 33476 times

The core in the +20% dynamic range image is not overexposing and, while more concentrated, does not appear star-like and retains a "fuzziness" compared to the stars.

Does that help?

[Mike in Rancho]

While trying to help figure out an issue, I had access to some BXT'd data, as well as the original stack (Thanks Steve!)

The file was linear but had been ABE'd (possibly with some settings problems), BXT'd, and PCC'd. So, I opened as white-balanced, still did a Wipe but zeroed out the aggressiveness slider, no-ROI OptiDev, Color, and Denoise.

For the original stack I ran a full ST processing, meaning Wipe and OptiDev, then the usual "carve down" sequence of Contrast but locality 25%, HDR with a few changes from default, Sharp, and SVD, followed by Color and Denoise. I tried to match the overall color balance and saturation the best I could eyeball it.

I also found a HST (infrared though) image by NASA, scaled and rotated that so I could flip between all three versions.

The data was about 3 hours of RASA 11 in Bortle 3 with a 2600MC. No binning. Images were displayed at 200% and I screenshot those and cropped to match. The NASA picture was scaled down to 58%, and I don't know if they did their own scaling in creating that image.

Setting aside the stars (across the full image BXT seemed to mush both a triple star system and a double star together a bit, and even in the Flame crop it's not resolving hidden stars), I don't necessarily see anything being made up, though I suppose that depends how you define it. But it does seem like BXT is shrinking feature details and giving them a fine-edged "clarity" that I'm not sure is reasonably supportable.

Or, is it possible I'm doing things wrong in ST such that I'm fattening details up or leaving clarity on the table? As a comparison point, I was mostly focusing on that little arrowhead shape to the left side. I also ran a ST workflow where I only applied SVD and took it up to 20 iterations. Also tried unsharp mask in Gimp, deconvolution in Siril, and even threw Registax wavelets at it. I'd have tried PI deconvolution but it seemed like way too much work. But regardless of whatever I used, the features all stayed more or less the same size, but with varying degrees of resolving. Only BXT results in those tiny fine-grained details of the same structures.

Well, except for HST infrared, which even scaled down is just off-the-charts sharp.

ST's full processing did bring out more detail overall (without the fine edges), but admittedly the BXT file was only stretched with no other enhancements applied.

In order, Original stack 200% crop, Stack with BXT 200% crop, NASA HST image 58% downscale.

Steve Orig FullST Flame crop 200pct.jpg (451.46 KiB) Viewed 33148 times

Steve BXT etc Flame crop 200pct.jpg (456.7 KiB) Viewed 33148 times

Flame Comp HST rot scale.jpg (465.32 KiB) Viewed 33148 times

The three files are pretty well aligned such that they can be overlaid and blinked/viewed, such as in Gimp layers.

Thoughts?

[admin]

Setting aside the stars (across the full image BXT seemed to mush both a triple star system and a double star together a bit, and even in the Flame crop it's not resolving hidden stars), I don't necessarily see anything being made up, though I suppose that depends how you define it.

But it does seem like BXT is shrinking feature details and giving them a fine-edged "clarity" that I'm not sure is reasonably supportable.

The "magically thinning structures" is indeed something I observed/mentioned early in the thread.
It is a clear example of overfitting an AI model, as the thinning of structures is "usually"/"often" what happens but - in reality - not always or necessarily;
As I mentioned there;

The Centaurus A image on that page definitely displays the "this-is-what-you-want-to-see, right?" magically thinning structures (which cannot be corroborated in closer up images.) so reminiscent of the Topaz "dog hair" filter. The choice of a galaxy to demonstrate the module's prowess is somewhat unfortunate, as they are objects where detail is much easier to resolve further in higher quality images (rather than, say a nebula in our own galaxy that becomes more diffuse at larger scales).

E.g. what should be structures that "fall apart" into smaller structures of diffuse dust when better resolved, with BXT become solid/defined thinning interconnected strands.

Or in other words, if you train an AI to turn everything into sharply defined features - even when the input image cannot possibly support this - you will get exactly that "wishful thinking" detail. Not every "ideal" result is a Hubble or JWST image. What is "ideal" is a function of the input data's characteristics. BXT's premise is fundamentally flawed if you are after an image that depicts reality. If you use an agnostic model based on physics/optics you get reality.

That there should be no "thinning" happening, can be clearly seen in this excerpt of a deeper image in the same spectrum by Adam Block;

thinning.png (225.46 KiB) Viewed 33119 times

(Source of excerpt: Adam Block/Mount Lemmon SkyCenter/University of Arizona)

Just like with numerous of examples of ChatGTP being "confidently incorrect" in its responses, so is BXT. It's what happens if you do not leave room for precisely measured uncertainty and demand an answer where the input cannot support the demanded precision of the answer.

An example of this is the area indicated;

bxt.png (464.25 KiB) Viewed 33123 times

BXT is synthesising (incorrect/non-existent) interlinked detail in an area that ST - rightfully - leaves as diffuse grain. (Note by the way that scaling up the ST crop, as done here, makes the grain stand out a lot more, as ST exploits noise grain retention for psychovisual purposes at <=100% scale).
That there should be no detail visible there, can be clearly seen in another excerpt of the deeper image by Adam Block;

AdamBlockFlame.png (252.36 KiB) Viewed 33123 times

(Source of excerpt: Adam Block/Mount Lemmon SkyCenter/University of Arizona)

At the same time BXT does not recover detail that should be readily recoverable by a true deconvolution routine.

My cautions stand; do not use BXT unless you are OK with made-up, re-interpreted detail. It is not deconvolution.

[Mike in Rancho]

Yes, perfect term, magically thinning structures. That should have sounded familiar. Sorry for making you repeat yourself!

I don't think I quite grasped it before, until I had some data to look at right in front of me, and able to see my own side-by-side plus now the new (better) ones you posted up.

Also, excellent new learning for me on what "deep" is, clarifying and resolving detail (if the resolution is there) but not altering the true, or at least true-to-the-data, spatial size of the features. If I have that right.

[Burly]

Now if anyone can train Ai to use startools id like to see how that works out theoretically could that be possible.

[Mike in Rancho]

Now if anyone can train Ai to use startools id like to see how that works out theoretically could that be possible.

Sure. And if not integrated as part of ST 1.20 beta, it could be a standalone but with some kind of porting and control capability, like STReplay.

Those constraints would solve some problems. Even if the neural net's decisions themselves could not be untangled, the final actions would all end up in ST logs, and thus traceable and reversible.

The question is maybe whether it would do any good, as compared to hard coded assistance algorithms or just plain end-user manual control.

And any automated processing will end up with proclivities and a "style," again whether AI or straight algorithm, and thus query whether that resultant style is something you actually like or want. The variance in targets and dataset quality is vast. Even ST's current carryovers of settings from one module to another, obviously based on intelligent assumptions, don't always hit the mark and require some human tweaking.

[dx_ron]

It will be interesting to see the thread that develops from this CN post: https://www.cloudynights.com/topic/8936 ... p=12974562

[Mike in Rancho]

It will be interesting to see the thread that develops from this CN post: https://www.cloudynights.com/topic/8936 ... p=12974562

Interesting. A little experiment confirming that Ivo was right oh so long ago.

EDIT: Well, I think it fizzled out. The OP there found a lesser BXT setting that results in only slightly more "clarity" than ESO's professional telescope, instead of way more, and is now happy with it.

Sigh.

[dx_ron]

BXT 2.0 turns every flawed imaging train into perfection - producing perfect round stars out of almost anything.

Actually, my real point in posting today is that Croman has put up a description of the math that BXT is attempting: https://www.rc-astro.com/the-mathemati ... erminator/

Not so useful to me, but hopefully useful for those less math-challenged.