StarTools 1.9 preview

[Mike in Rancho]

edit - oops.

[Carles]

but in some bright stars creats this artifact

Hi Carles,

this looks a bit like the artefacts we discussed a few days or weeks ago:
viewtopic.php?f=4&t=2821&p=14291&hilit=centroid#p14291
(There have been some more posts regarding this issue in the "StarTools 1.9 preview" and other threads.)
Have you switched on the 'PSF Resampling' 'Centroid Tracking' option and/or checked for dud PSF samples?

Best regards, Dietmar.

Yes, of course I tried, and it doesn't solve that artifact.

[Mike in Rancho]

Yeah Carles, the new SVD and deringing regime can be a bit finicky, dare I say tempermental?

Your shadowing is really shifted though, which kind of points towards a bad sample. Were you loading PSF's at all?

I actually ran across that little snafu tonight after a restack that was very very similar, but had actually used a slightly different reference frame. When I loaded the PSF samples txt from previous processing, the selected sample boxes aren't recentered on their stars, and so I had a slight wobble to my ringing and shadowing that looked similar to your image.

I'm still struggling myself. Sometimes you can go back and alter early workflow, such as DAF, IFD, and the OptiDev, but for one, that really starts driving processing into one narrow path just to get SVD to work, and two that doesn't always solve things anyway.

It still strikes me that 536 deringing might be nice as an option B, though maybe that algorithm is obsolete due to the other changes in SVD. I probably need to practice and experiment more, but right now I'm spending a lot of time making custom masks for shrink-dering, and since that doesn't always fully repair the flaws or has sufficient color match, custom masks for layer blurring as well.

Here's one I had earlier, one little crop region that displayed a very heavy dark ring (SVD just may not like my Newt), nearby stars that seem basically unaffected, and then a couple (one in particular, lower and 3/4 right) with almost a reversal or ghosting from the deringing.

551Ringydingy.jpg (465.08 KiB) Viewed 36960 times

On another, and probably easier, note, in Color the Before-After toggle button only shows Before-Before when in MaxRGB screen. It still toggles correctly, just the button doesn't change.

[alacant]

1.9.551 Ubuntu 22.04 i5 8Gb

Hi everyone

SVDecon sampling takes -quite a bit- longer than previous versions and you can''t 'mouse ahead'; you have to wait for the star to be selected before moving on.
Anyone?

Cheers and clear skies,
Steve

[Carles]

Yeah Carles, the new SVD and deringing regime can be a bit finicky, dare I say tempermental?

Your shadowing is really shifted though, which kind of points towards a bad sample. Were you loading PSF's at all?

I actually ran across that little snafu tonight after a restack that was very very similar, but had actually used a slightly different reference frame. When I loaded the PSF samples txt from previous processing, the selected sample boxes aren't recentered on their stars, and so I had a slight wobble to my ringing and shadowing that looked similar to your image.

I'm still struggling myself. Sometimes you can go back and alter early workflow, such as DAF, IFD, and the OptiDev, but for one, that really starts driving processing into one narrow path just to get SVD to work, and two that doesn't always solve things anyway.

It still strikes me that 536 deringing might be nice as an option B, though maybe that algorithm is obsolete due to the other changes in SVD. I probably need to practice and experiment more, but right now I'm spending a lot of time making custom masks for shrink-dering, and since that doesn't always fully repair the flaws or has sufficient color match, custom masks for layer blurring as well.

Here's one I had earlier, one little crop region that displayed a very heavy dark ring (SVD just may not like my Newt), nearby stars that seem basically unaffected, and then a couple (one in particular, lower and 3/4 right) with almost a reversal or ghosting from the deringing.


551Ringydingy.jpg



On another, and probably easier, note, in Color the Before-After toggle button only shows Before-Before when in MaxRGB screen. It still toggles correctly, just the button doesn't change.

Yes, I was selecting PSF's, and tried different ones, different sizes (unselecting previous then selecting new ones )
I miss the control capability of 1.8, vs 1.9. the controlos on 1.9 SVD for deringing are not enough for me and my datasets.
will keep trying new 1.9s but for now, will keep processing generally with 1,8.

1.9.551 Ubuntu 22.04 i5 8Gb

Hi everyone

SVDecon sampling takes -quite a bit- longer than previous versions and you can''t 'mouse ahead'; you have to wait for the star to be selected before moving on.
Anyone?

Cheers and clear skies,
Steve

Same here with W10 , Ryzen 7 5800h , RTX3070, 64Gb ram. latest sdv decon takes quite a bit, gets sort of "frozen" to the point of not being able to select a preview rectangle either for few seconds, and also each sample selecting makes this happening.

[dx_ron]

I have an M51 stack where I let a couple of bright stars saturate. In SVD they are not included in the mask (and there is no way to edit the mask - though perhaps mask editing is unwise?).

The saturated stars end up with substantial ringing, and the ringing seems totally unaffected by the Deringing slider.

Before:

before.jpg (464.29 KiB) Viewed 36917 times

Mask:

mask.jpg (295.18 KiB) Viewed 36917 times

After:

after.jpg (417.41 KiB) Viewed 36917 times

[Mike in Rancho]

Yeah Ron you are probably on to something, and Ivo has recently mentioned the fully saturated core issue.

I wonder if it's that alone, or would also have anything to do with the sample sizing? Not like we would choose that sample anyway, since it would probably have a red core.

I guess the question becomes, what does SVD do when it runs into a saturated circle of a blown star, when applying the regional PSF deconvoluton? Well, makes rings I guess.

I need to survey some of my stacks, but I'm not sure with what. Something like the graph that PHD2 shows of the guide star would be good. In PI I've found a 3D plot it can make, though a bit awkward as instead of doing it on a selection area you have to crop down an image. But I might run that on some cropped blown and not-blown stars, both linear and autostretched, so see if that might show me a PSF shape.

I think one problem is, sacrificing a handful of star cores to saturation is extremely common. In fact on CN (both BDSI and EDSI I think) the general guidance for exposing is to have NINA max pixels somewhere in the low 100's. This is dependent on just how many bright stars are in the field, of course, as well as if any are mega stars, like Arcturus or Vega.

I usually take 30s L and 60s RGB, but I could probably take 5 or 10 seconds L's and still blow some central cores, so it's just not feasible. The 2600MM has a decent full well I think, though not really expanded like say the Player One Poseidon. It would be nice to have that.

This also makes me wonder if I could end up with mangled cores (therefore troublesome to SVD) from L+Synth L, since my color filters will not necessarily match up as far as core saturation. Usually in NINA I see the RGB maxed pixels to be a good bit less than the L, again at 30s/60s. Maybe 30s/90s would be a closer match?

I don't have the computing power to drop to 10 second subs, and just think of how many subs that would be, plus all the gigabytes it would take. And stacking would take a week.

[dx_ron]

In 1.8, using the Alt-Stars Auto Mask, you can click around with Flood Fill to include the saturated stars that the auto-mask skipped. I obviously would not choose one of those as a sample, but it makes a big difference in how the star comes out after deconvolution. The big issue with 1.9 is that those stars remain unchanged no matter where I set the deringing slider.

I should start making notes about how many saturated pixels I allow. I think it was probable 2-300 for this (Low-Conversion Gain, 240s at f/7 - I think that's equivalent to ZWO's gain 0). Nevertheless, it appears that saturating just a pixel or four at a star core does not lead to this - it appears to be only the extra bright stars. The two stars hardest hit in this image are mag 7. I've started shooting for fewer than 100 saturated pixels per sub, but a bright enough star will land in the same boat.

I think ZWO is now the lone holdout to not let users choose the "extended full well mode". It is "even lower conversion gain" - maybe ZWO feel boxed in by their choice to already call something gain 0. Both QHY and ZWO amuse me with how they act like they invented something new when they eventually decide to implement something that already comes with the chip (meaning that it was "invented" by Sony). QHY recently did that with "Low-Noise Readout" mode on the IMX571, putting out a press release that made it sound like they had discovered a new CMOS technology - while Touptek had enabled Low-Noise mode from the start and never really said much about it.

[admin]

I just released a new alpha with more control over the de-ringing parameters (as well as a substantial improvement to processing speed when using intra-iteration resampling). I think you have all helped me realise that there are just too many ways datasets can be compromised that require more hands-on control, e.g. the non-linear saturation-approaching case, as well as imperfect sampling). I have separated-out the single parameter into three parameters.

Let me know if that solves your particular issues - I have had good success getting rid of ringing in most of the cases you have been bringing up and demonstrating...

[Mike in Rancho]

Thanks for the work, Ivo.

The controls do allow customization which results in improvement over 551. Plus, the speed!

Experimenting and working on some comparison side-by-sides, trying different things, so will post those up once finished. Are IFD and DAF etc still incorporated or is that out?

Once I get a 552 final as good as I can I'll have some crops to put up. So far I've run it against 551, 536, and 8-527 with both starfishy and um, sand dollar sampling. Gosh, it's easy to forget how much 1.9 is improved and speedier than 1.8 overall. I also forgot my brutal star shapes from back then. All the 1.9's just handle my Newt star cores so much better.