Beware of star fixing tools

[Adam J]

12 hours ago, Scott Badger said:

I don't know of any processing tool that gives you perfect results, and no one, especially the developer, is claiming that BlurX does. And like any software in continual development, there may be bugs that need fixing, and according to a CL thread, it appears the latest releases of BX does have a problem, especially when used in the auto psf mode. When not broken and used properly, though, it can do deconvolution (the algorithms it's actually using) better than we can mostly because it can work with millions of parameters, tens of millions even, where we can only handle a few. But in the end it's still up to the operator to look at the result of any process, assess it, and decide whether it's an improvement. You can blindly trust the result of any processing software. Or not.

It *is* true that the nature of a trained neural network tool means a final result can't be 'predicted', or even back-engineered i.e. no matter how well you know the code and the training, you can't know exactly what leads/led to a particular result. But for our work, I don't see why that matters. I couldn't follow the math, let alone the code, in most of the processes we use, 'AI' or not. For research purposes, that opaqueness *is* a problem, but again, Russ Croman is the first to say that BlurX or any AI software shouldn't be used for that purpose. That said, professional astronomers have recently gained a better understanding of the M87 black hole by using AI to sharpen the image.

There is also the DIY vs DFY (done for you) divide, but for me it's like woodworking, there are some things I like to do with hand-tools, and others where power tools and guides are my choice. Positioning this as all or nothing, if you use BX you might as well just go out and get a Stellina and be done with it, is a straw man argument. It would be just as fair for an astrophotographer that uses film and manual guiding to level the same charge at anyone using electronics......

Arguments regarding image manipulation in general get muddy really fast, and the use of AI has little to do with it. Generally speaking, there's corrective and aesthetic; sharpening stars and fixing shapes is corrective, reducing the number of stars is aesthetic. That's pretty clear. But what about background and the use of tools like DBE and noise reduction, where's the line between correction and aesthetic improvement there? Or how we stretch images and insert contrast in a non-linear way, especially with tools like GHS? And then there's color...... I really don't see how any line drawn is anything but personal, or why anyone should feel compelled to process their images in a particular way. You can like, or not like, my images for all sorts of reasons, and the world goes on.

Cheers,
Scott

 

12 hours ago, Scott Badger said:

I don't know of any processing tool that gives you perfect results, and no one, especially the developer, is claiming that BlurX does. And like any software in continual development, there may be bugs that need fixing, and according to a CL thread, it appears the latest releases of BX does have a problem, especially when used in the auto psf mode. When not broken and used properly, though, it can do deconvolution (the algorithms it's actually using) better than we can mostly because it can work with millions of parameters, tens of millions even, where we can only handle a few. But in the end it's still up to the operator to look at the result of any process, assess it, and decide whether it's an improvement. You can blindly trust the result of any processing software. Or not.

It *is* true that the nature of a trained neural network tool means a final result can't be 'predicted', or even back-engineered i.e. no matter how well you know the code and the training, you can't know exactly what leads/led to a particular result. But for our work, I don't see why that matters. I couldn't follow the math, let alone the code, in most of the processes we use, 'AI' or not. For research purposes, that opaqueness *is* a problem, but again, Russ Croman is the first to say that BlurX or any AI software shouldn't be used for that purpose. That said, professional astronomers have recently gained a better understanding of the M87 black hole by using AI to sharpen the image.

There is also the DIY vs DFY (done for you) divide, but for me it's like woodworking, there are some things I like to do with hand-tools, and others where power tools and guides are my choice. Positioning this as all or nothing, if you use BX you might as well just go out and get a Stellina and be done with it, is a straw man argument. It would be just as fair for an astrophotographer that uses film and manual guiding to level the same charge at anyone using electronics......

Arguments regarding image manipulation in general get muddy really fast, and the use of AI has little to do with it. Generally speaking, there's corrective and aesthetic; sharpening stars and fixing shapes is corrective, reducing the number of stars is aesthetic. That's pretty clear. But what about background and the use of tools like DBE and noise reduction, where's the line between correction and aesthetic improvement there? Or how we stretch images and insert contrast in a non-linear way, especially with tools like GHS? And then there's color...... I really don't see how any line drawn is anything but personal, or why anyone should feel compelled to process their images in a particular way. You can like, or not like, my images for all sorts of reasons, and the world goes on.

Cheers,
Scott

It only matters to me for three reasons, 1) If the result looks fake 2) If the result unintentionally adds elements that did not appear at all in the original image. 3) If the result removes elements from the original image that you don't want to loose. 

For me the example above covers all three. 

1) I find the stars too perfect to be believable. 

2) A weird approximation of a planetary nebula was added. 

3} In adding that nebula like object a star was totally removed from the image as opposed to repaired. 

It could well be setting that the op has selected causing these issues. But that's another reason to take care

Now does that mean I would not use it myself. No I would use it. It means that you need to keep a very close eye on what it's doing. For me the biggest sin is adding anything not real to the image as it removes it too close to art and to far from the science sides of the hobby. 

To put a finger on it I believe it's sometimes treating bright linear structures as stars. Another one I have seen is it creating little blobs ob diffractions spikes. All in all it doesn't matter, but it makes me think more is going on than just the claimed AI administered deconvolution. So I would be checking my images. 

It's just an opinion don't be offended by it. 

Adam

Edited July 31, 2023 by Adam J

[windjammer]

I don't think it is a matter of winning or losing - no one should mind when fair points are made on processing.  And it helps people like me pick up on where we go off-track: my stars are nearly always crap so the info on star tools was very interesting (I might even buy it). False detail/colour and over sharpening generally is another issue for me - I never get the same answer twice if I repeat the processing.  So any discussion on how to avoid artefacts is v. interesting - more useful even than looking at more images: generally no-one posts enough info on how they did their processing.

I had a look at my own Veil after your post - compared before and after BXT on the master lights and didn't see anything obvious.  So I think you should go ahead and post, you would have a willing audience of least one!

Simon

[Scott Badger]

10 hours ago, Adam J said:

 

It only matters to me for three reasons, 1) If the result looks fake 2) If the result unintentionally adds elements that did not appear at all in the original image. 3) If the result removes elements from the original image that you don't want to loose. 

For me the example above covers all three. 

1) I find the stars too perfect to be believable. 

2) A weird approximation of a planetary nebula was added. 

3} In adding that nebula like object a star was totally removed from the image as opposed to repaired. 

It could well be setting that the op has selected causing these issues. But that's another reason to take care

Now does that mean I would not use it myself. No I would use it. It means that you need to keep a very close eye on what it's doing. For me the biggest sin is adding anything not real to the image as it removes it too close to art and to far from the science sides of the hobby. 

To put a finger on it I believe it's sometimes treating bright linear structures as stars. Another one I have seen is it creating little blobs ob diffractions spikes. All in all it doesn't matter, but it makes me think more is going on than just the claimed AI administered deconvolution. So I would be checking my images. 

It's just an opinion don't be offended by it. 

Adam

No offense taken and enjoying the discussion. I don't disagree with anything you say above, other than your suspicion that more than just deconvolution is involved. I can't prove it, but Russ Croman has been very forward with information about the tools he's created, and AI in general, and for me at least, he's established enough credibility that I'm willing to to take him at his word unless or until there's clear evidence that something else is going on. That said, there's certainly no question that BX, like any other decon tool, can create artefacts if the settings aren't optimal and sometimes, depending on the quality of the data, it won't have much effect at all. Like other tools still in development, bugs can also come up, and as I mentioned before, I think the OP's image may be an example of that. I've not seen anything like it with my use of BX, and there was a similar report in another thread (CL or the PI forum) from someone who had just gotten the new version (I haven't updated mine yet).

To your examples, I've not seen it turn structures into stars, but I have seen the reverse; stars strung into a filament like structure. This can happen (as it happened to me with Andromeda) when the stars are small and part of a larger scale structure, like a galaxy. Using the manual psf setting and enabling "Nonstellar then Stellar" solved that particular issue, but in the end you sometimes just have to back off on the amount setting (or use a smaller psf), even if you aren't getting as much improvement as you were hoping for. Anyhow, it's no different than any other tool in that artefacts are possible and like you said, it's up to the operator to assess the results and try again if necessary. Something to note as well is that BX uses a tiled approach (part of what it can do that we can't with traditional tools), so depending on the number of stars in a particular tile, or the quality of stars in that part of the image, the sharpening effect and/or artefacts created can vary across the image. That's why the manual psf mode should always be used, and maybe why it can appear that an artefact is something 'added' to the image (i.e no direct/mathematical relationship with the image data) when it occurs in just one area as opposed to throughout the image like we'd see with a traditional decon tool.

Cheers,
Scott

[Adam J]

5 hours ago, Scott Badger said:

No offense taken and enjoying the discussion. I don't disagree with anything you say above, other than your suspicion that more than just deconvolution is involved. I can't prove it, but Russ Croman has been very forward with information about the tools he's created, and AI in general, and for me at least, he's established enough credibility that I'm willing to to take him at his word unless or until there's clear evidence that something else is going on. That said, there's certainly no question that BX, like any other decon tool, can create artefacts if the settings aren't optimal and sometimes, depending on the quality of the data, it won't have much effect at all. Like other tools still in development, bugs can also come up, and as I mentioned before, I think the OP's image may be an example of that. I've not seen anything like it with my use of BX, and there was a similar report in another thread (CL or the PI forum) from someone who had just gotten the new version (I haven't updated mine yet).

To your examples, I've not seen it turn structures into stars, but I have seen the reverse; stars strung into a filament like structure. This can happen (as it happened to me with Andromeda) when the stars are small and part of a larger scale structure, like a galaxy. Using the manual psf setting and enabling "Nonstellar then Stellar" solved that particular issue, but in the end you sometimes just have to back off on the amount setting (or use a smaller psf), even if you aren't getting as much improvement as you were hoping for. Anyhow, it's no different than any other tool in that artefacts are possible and like you said, it's up to the operator to assess the results and try again if necessary. Something to note as well is that BX uses a tiled approach (part of what it can do that we can't with traditional tools), so depending on the number of stars in a particular tile, or the quality of stars in that part of the image, the sharpening effect and/or artefacts created can vary across the image. That's why the manual psf mode should always be used, and maybe why it can appear that an artefact is something 'added' to the image (i.e no direct/mathematical relationship with the image data) when it occurs in just one area as opposed to throughout the image like we'd see with a traditional decon tool.

Cheers,
Scott

Interesting thought occurs to me in that turning the settings up to the max, while not visually preasing, might produce extreams of results that give a clue as to the underlying process or direction being used by the AI.

Adam

[Scott Badger]

1 hour ago, Adam J said:

Interesting thought occurs to me in that turning the settings up to the max, while not visually preasing, might produce extreams of results that give a clue as to the underlying process or direction being used by the AI.

Adam

Possibly, but depends on the data I think.....  My seeing is generally poor to terrible and I'm very oversampled at 0.33"/pixel, so it's rare that my psf is less than the 8 pixel maximum, plus I generally run BX at the default 0.9 sharpening amount, so can't really push it much further. I've also found, though I haven't heard others make this comment, that when I've tried using BX on data with poor SNR (some of my first images with little exposure time and taken with a dslr), it didn't really go awry, it just didn't do much at all. 

Something else to note, and not sure how much difference it makes, but according to RC, BlurX wasn't trained on stars any bigger than 8 pixels. For most that should be fine, but if your seeing is poor and you're imaging at a relatively long focal length (like me), the results may not be optimal.

BTW, I would encourage the OP to send the problem image to RC. He's very responsive to any inquiries and I'm sure he'd like to see an issue like that.

Cheers,
Scott

Edited July 31, 2023 by Scott Badger

[tomato]

19 hours ago, Scott Badger said:

 

BTW, I would encourage the OP to send the problem image to RC. He's very responsive to any inquiries and I'm sure he'd like to see an issue like that.

 

Thanks for the suggestion, I never thought guys like RC would have time to answer individual queries, but I will email him and let folks know on this thread what happens.

[Scott Badger]

Yes please. It would be great to hear any feedback he gives

Also, he’s done a number of interviews/presentations with TAIC, Adam Block, VisibleDark, etc. where he discusses his tools and AI as applied to AP that are pretty interesting.

Cheers,
Scott

[tomato]

In preparing a proposed email to RC Astro I have gone back and reproduced the star correction workflow which created the PN artefact from Epsilon Cygni. It turns out that BXT played no part in creating the artefact. All extremes of settings in BXT failed to produce the effect, the problem arose when I took it into StarTools after BXT and created an auto star mask. The default settings masked the spikes on Epsilon Cygni, but left some of the star untouched:

When applying the star repair tool, redistribute, core is average location, with all other settings on default, the PN is created:

Manually filling the mask eliminates the problem. So my lesson learnt is don't assume the auto default settings will always have the desired effect, and apologies to Russell Croman, for putting BXT under suspicion.

 

[windjammer]

Well Mr T - I am amazed that you produced a great image from the starting point you had.  Well done.  The raw star images were terrible! I would have given up.

So, after all that, we have a bit of a bodge around a very bright and over exposed star (who knew!) and not much else *demonstrated* as artefact or spurious ?

Amazing.

Simon

[tomato]

Thanks, I did say the star shapes in the raw data on the LHS were horrendous, but the subject of the image was much better quality so I decided to proceed even though this meant hitting the stars with both processing barrels so to speak. Their authenticity suffered as a result but I’m still glad that I didn’t consign the data to the bin.

[Scott Badger]

Astrophotography is a sea of red herrings….

Cheers,

Scott