Astro to Astro-art Continuum, 2024 Edition

[jlh]

Thank you guys for the excellent discussion about sharpening.
From reading Professor Morrison's article, it is accurate to conclude that Startool's wavelet sharpening is a more data-driven tool than other so-called "accutance" sharpening algorithms? That is, it is essentially a correction using the intrinsic data available and not just more or less an enhancement of the image.

Thanks again. Following these discussions (or at least trying to) makes the use of Startools to achieve truly accurate images more understandable and more enjoyable to use.

Jeff

[decay]

While likely broad categories compared to how ST handles things in a more complex and astro-specific manner, there were a few Wikipedia entries that cover stuff like the scaling involved in wavelets.

Yes Mike, I already saw all these Wikipedia entries. But we need a lot of background knowledge and mathematics to understand all this. My knowledge is limited; I learned a lot about signal processing, but this is a long time ago. Or maybe in a previous live? And I haven’t used it since then. And it was only about 1D processing (audio, radio signal transmission etc.). (But now since Ivo pointed out the 1D equivalents we may break things down …)

So I tried to find something easier to understand about Wavelet Sharpening, but all I found was what I already posted in Freddy’s tadpoles thread:
https://photo.stackexchange.com/questio ... n-registax

(And the answer explains (at least a bit) that and how the ‘blur/subtract/add edge detections’ you asked for do work in this case. )

But maybe we can think of it a bit easier: Ivo told us (BTW: thanks!) to think of Wavelet Sharpening as of using a ‘graphic equalizer’. As we know, every signal can be decomposed into a base frequency and it’s harmonic waves. (To be correct, this is simplified and true only for periodic signals.) The way back, if we know the frequencies and their amplitudes (and phase angles), we can compose them and as result we get back the original signal. But what if we change the amplitude values of some frequencies in between? This is what a graphic equalizer does. So you can decide to pronounce human voices and dial in frequencies between 100 Hz and 1 kHz for example. All the other noise remains as loud as it is.

So Wavelet Sharpening decomposes the image into different spatial frequencies (scales) and using the sliders we can change the contribution of this frequency ranges independent of each other.

Dietmar.

[decay]

Hi Jeff, glad to hear that you are still following I suppose, we are making a lot of noise here, to use Steve's phrase

it is accurate to conclude that Startool's wavelet sharpening is a more data-driven tool than other so-called "accutance" sharpening algorithms? That is, it is essentially a correction using the intrinsic data available and not just more or less an enhancement of the image.

I'm not sure. Yes, ST's Sharp module uses the intrinsic data, but I suppose, the described acutance sharpening algorithms do just the same. And in both cases I think, 'enhancement' is an applicable term. And 'correction' possibly not. Unless we are 'undoing' something that happened to our data. Let's say the high frequencies of the signal were dampened on the way down to us. Now we can use wavelet sharpening to amplify these high frequency components exactly up to the intensity they had before. In this case it would be a 'correction' or 'restorative' to use Ivo's term. But of course, you would have to know how much was dampened and how much to amplify ...

Of course, ST's wavelet sharpening allows for the 'sharpening' of lower frequency components as well (as described above).

Dietmar.

[Mike in Rancho]

• Wavelet sharpening is about isolating and manipulating specific frequencies (like a graphic equalizer).

Hi Ivo!

Well, I had a graphic equalizer once. Once.

And while I did toy around with electronics a little bit in my youth, I was never the type to have, ya know, an oscilloscope. Also, historically, my soldering has always been functional but not pretty. At all.

So even from a 1D or audio stream standpoint, there's a lot of Greek there for me.

(What is UM?)

Unsharp Mask. Sorry for the acronym.

We should think about what we want to know ... and what to do first ... I'm confused

I will check your app out. I think what we want is a good sample image, or set of sample images. Yes synthetic or even terrestrial images may be useful as a tutorial sample, but maybe also we can find an astro image with some features that would be good for showing things (and also have stars so we can run SVD on it)?

Last night I tried to do more reading on wavelets, and mostly some detailed write-ups on how R6 is probably functioning. Then I played with some finished images and ran Gimp default settings UM on it, then opened in PI to run the 3D graph generator on a ROI of the Flame for before and after. Indeed it just seemed to raise the mountains and ridges and perhaps deepen the canyons of the 3D graph.

Now...I think that is expected behavior but adds another wrinkle to my thinking. In addition to steepening the transitions, the data also seemed to be expanding its range on the Z axis.

That is, it is essentially a correction using the intrinsic data available and not just more or less an enhancement of the image.

Maye not quite, yes I accept Sharpen is in the intrinsic category, but also enhancing, along with Contrast and HDR, and my initial query was trying to divvy those up even more, thinking Sharpen to be further into enhancing and away from data fidelity than those other two. Which Dietmar answered early on as maybe they really should all be considered about the same.

Which is fine, and rehabilitates sharpening a little bit in my mind. At least for wavelets and ST sharpen, but I'm still not very confident in UM.

Last night I fiddled with Sharpen some, and I believe I found one of my perceptual problems with it. Default is large structure, but also has all wavelet sliders maxed out at 100%. Well, #5 and sure #4 are pretty large scale, and that (leaving things default) is why I have mostly seen what looks mostly like a simple brightening of bright stuff. And in a lot of cases I didn't find terribly useful, or like anything I would see in R6 doing planets.

Of course R6 starts with everything at zero.

Well, duh! I'm an idiot.

It seems to me that, especially with large selected in screen 1, for a lot of our datasets it's almost mandatory to zero out the #5 and #4 scale sliders, and maybe even #3 as well. And with scaling more appropriate to the portions of the target that are of interest, the strength can be jacked up as well.

I'm such a bonehead.

[admin]

Wavelet Sharpening in StarTools is indeed "more" data driven than a regular wavelet sharpener, in that it uses more statistics and analysis of the data to achieve two - for AP generally - desirable things;

• it refrains from boosting areas that don't have enough signal (uses Tracking's signal evolution stats for that), avoiding noise becoming prevalent
• it allows for giving priority to higher frequencies (smaller scales) if multiple bands/scales are competing

Or in terms of audio;

• it doesn't turn up the volume for a frequency band if there is no signal there (e.g. only noise would be audible)
• it doesn't turn up the volume for the bass (lower frequencies) if the lead/melody (higher frequencies) is also playing; when everything is loud, nothing stands out

In general, wavelet sharpening (and UM for that matter) do not add new data (unless clipping). The only transform the data and are therefore "data driven" as well if you will.

[decay]

Hi Ivo, thank you for these additional comments, and for 'translating' them into their audio equivalents.

I wonder if the audio analogy has it's limit when we look at the default settings of the scale sliders. By default, details of all scales are sharpened (100%). In terms of audio we would push up all sliders of our graphic equalizer. I suppose, the result in this case is just that volume is turned up, i.e. the amplitude of the signal is higher / multiplied by a constant factor.

In terms of image processing 'more volume' means 'brighter' I suppose? But that is not the case, instead it's something like more contrast ... ?

Best regards, Dietmar.

[decay]

Hi Mike,

we should have taken a look at Guy's User Notes beforehand:
viewtopic.php?f=12&t=1777

Much of what we discussed is explained there in excellent detail ...

And be sure to check out the mentioned link:
http://keithwiley.com/astroPhotography/ ... ning.shtml

And additionally there's a lot of infomation regarding the scales.

And ... and ... and ... I'm feeling a bit foolish

Best regards, Dietmar.

[Mike in Rancho]

Hi Dietmar,

Yes that extra link in the notes is pretty good, and has some nice graphs to show things working.

I'm still a bit flabbergasted that blur subtracts accomplish this, and while I get that it is intrinsic, I'm still trying to figure how narrowing the transitions and increasing amplitudes doesn't "add" structure or change spatial sizes. I must mull it over more.

I'm not sure volume or even equalizer levers is a direct analogy. The graphs in that link seemed to show increases away from the baseline, on either side, in order to achieve that sharpening/acutance. The volume knob or slamming all the EQ levers up would just raise everything, no?

But then again, like I said it now seems to me that if all the Sharpen scale sliders are at 100%, lower scale detail enhancement gets lost in the wash. So...maybe.

I'll have to experiment some more and perhaps with bright/dark not set to 50/50.

[admin]

Hi Ivo, thank you for these additional comments, and for 'translating' them into their audio equivalents.

I wonder if the audio analogy has it's limit when we look at the default settings of the scale sliders. By default, details of all scales are sharpened (100%). In terms of audio we would push up all sliders of our graphic equalizer. I suppose, the result in this case is just that volume is turned up, i.e. the amplitude of the signal is higher / multiplied by a constant factor.

In terms of image processing 'more volume' means 'brighter' I suppose? But that is not the case, instead it's something like more contrast ... ?

Best regards, Dietmar.

Processing in the audio domain is identical, but greatly simplified due to there being one fewer dimension to take into account. It is precisely what you intuitively did with your helpful tool; remove one dimension to reduce complexity and improve understanding.

Keep in mind there is always the "scale 0" residual(s) the boundaries beyond which the scales do not operate (for an image the "very" blurred image, for audio low frequency rumble and ultra-high frequency noise/overtones). The amplitude of scales undulate around that. For wavelet sharpening you control the amplitude of those undulations; whether those undulations are visual (brightness changes, aka acutance) or audible (volume changes) is immaterial.

So, in terms of your tool, it would be helpful to calculate the average brightness for the entire cross-section and subtract that average from what is displayed. This then shows the undulation in relation to the baseline, which is much more helpful in terms of measuring how humans would perceive the brightness changes/acutance (not the level of brightness by itself, which means very little to human eyes/brains).

This notion of the need for brightness changes to make things appear sharper through establishing (and enhancing) those undulations, also neatly highlights the difference between a sharpening tool (whose purpose it is to enhance such undulations) and a deconvolution tool (whose purpose it is to recover correct brightness levels for every sample/pixel, but is not occupied with enhancing undulations/acutance to make them more visible for humans.

Establishing correct brightness levels = data recovery; the realm of science and (mostly) objectivity.
Establishing acutance/emphasis through modifying brightness level undulation = enhancement; the realm of artistic choice and (mostly) subjectivity.

In neither cases is data invented from nothing though, which is important - neither cases necessarily invalidate the veracity of what is being shown.

[Mike in Rancho]

Establishing correct brightness levels = data recovery; the realm of science and (mostly) objectivity.
Establishing acutance/emphasis through modifying brightness level undulation = enhancement; the realm of artistic choice and (mostly) subjectivity.

In neither cases is data invented from nothing though, which is important - neither cases necessarily invalidate the veracity of what is being shown.

Thanks, Ivo.

Am I just chasing my tail then, trying to figure out which tools are "more" altering than others (such as Contrast or HDR vs Sharp)? I mean other than the intrinsic/extrinsic difference. Once you are into enhancing, it's either to amp up something so the human eye can perceive it, or a matter of imager's choice about what features to reveal or boost.

Correct! Just play around a bit and you will see ...

Dietmar I finally downloaded, unquarantined, and played with the brightness app.

One thing I quickly found out is that it requires the Windows display scaling to be at 100%, otherwise all wonky and it's reading stuff in places you don't expect it to.

I then made some simple sample images in Gimp, and I used a 50% background level in order to see how things diverged up/down after enhancement. I used a fuzzy paintbrush to draw both black and white lines as my sharpening targets. Gimp's built-in UM indeed worked as expected based on the reading we did, though it seemed slanted towards a deeper darkening than brightening.

I screenshot each stage with the Viewer in place so I could flip layers when pasted into a new image. I didn't go make a flashing gif, but...I could!

Before that I tested it using the "manual" method of UM via layer operations -- original + (original - blurred). Interestingly, while it worked, the Viewer showed that it was only brightening the light side of the transitions, not darkening the dark side. So, there must be another step involved. Or probably lots of steps, since the blur-subtract-add is very simplified I'm sure.

I also tried to fool it by drawing a fork in my line and bringing them near to each other, emulating fine structure in close proximity -- thinking the blur would overlap (it did) and ruin the sharpening or create incorrect information (it didn't). So, I gauss that blurring is smarter than I thought it was.