Thank you all for the feeback!
The TIFF file was a debugging remnant. I removed it for the latest upload; alpha "lucky" 13.
You can now find it in the download section.
I feel I'm finally nearing the end of this development cycle.
Mike in Rancho wrote: ↑Sun May 14, 2023 7:46 pm
1. First go through I got a mess of a sampling mask. Few samples, a lot of stars turned red, and what white outlines existed were broken up. I tried selecting one or two but the result was no good, either missing the centroid and/or shifting whole objects by many pixels. I restored, restretched, and tried again, at which point SVD just crashed out. I could not later replicate any of this. Maybe some errant click got me there? I took screenshots but if I never see this again I'll just chalk it up to some one-off aberration. Subsequent runs were much more normal.
Did you happen to have another version of StarTools running in parallel? It's usually
a clash of tracking files.
2. There are some blue pixels within some white outlines. Any meaning?
They do have a meaning indeed - I'm still deciding on this visualization. They show a couple of different quality aspects of the sample. Usually, samples with pure white borders are the best ones.
3. Am I noticing a difference in results depending on how I process the data on the way to SVD, or am I just seeing things?
Aside from Tracking of course having an influence, if you mean in terms of the samples it provides, then - for challenging datasets - it indeed it now takes into account your earlier estimates of noise that you made in other modules. E.g. if you used a high Dark Anomaly filter settings in Wipe or AutoDev, it will take this into account as an extra cue when trying to discern between stars and correlated background noise (because DAF settings target the same type of noise).
4. Deringing detect seems to act in the opposite manner from what I would expect, once one gets above 15% or so. Thus, I seemed to get a better deringing appearance when I dropped from the default 50% down to say 20. If I tried up to 70, ringing increased. Turning detect completely off causes kind of a harsh mess, but from 1 to again maybe 15 it seems to dering, before going backwards.
Indeed - it's definitley not a "strength" parameter any more. In my earlier post, I mentioned the nature of the slider has changed; it is balancing a number of different parameters. For challenging datasets (noisy and/or undersampled), experimentation on what works best may be in order.
In all cases, you should be able to achieve superior results in 550+ vs 536 (particularly in the visibility of darker structural detail), and in all cases touching the Dering Detect parameter on 536 will make darker detail/structures harder to see, while on 550+ the de-ringing will just look "different" but should impact structure much less.
5. Is deringing acting differently on my stars that end up outline sampled versus those that aren't? Maybe those are just bigger brighter stars, but they do ring more.
De-ringing targets stars (and structures) beyond the sample outlines. They are decoupled.
The approach to de-ringing in StarTools two-pronged;
- Detect and prevent ringing during deconvolution; this is the preferred cause of action (prevention!) and has yielded things like PSF Resampling. It involves actively adjusting the PSF during deconvolution, as well as estimating (and pulling back on) areas that may start to become "compromised".
- Detect and "repair" ringing after deconvolution; this is the "catch-all" fallback and the traditional way of countering ringing artifacts. It is more about restoring pixels that are deemed "too far gone" to their non-deconvolved state.
Good, active PSF estimation is key to preventing ringing. Gracefully/intelligently dealing with singularities (e.g. overexposed ares) and destabilisation is a close second. Cleaning up any remnants is third, but can still make or break an image.
6. In general with those larger stars, I seem to be getting a bigger diameter central white core than in 536. And in fact they seem less pinpointed than just from the OptiDev and pre-SVD.
I just made a slight tweak in alpha "lucky" 13 that reduces sizes a little.
Granted, this may just be the redistribution of energy inward, and something I need to get used to and/or use increase dynamic range to see what that does. I know you said 536 had some sort of flaw, but I still think it gives a better overall deringed appearance across the board.
I have been testing against 536 with various datasets, with various (but identical) samples and settings (including intentionally sub-optimal or very aggressive), and have consistently found quite the opposite; you should see much less
ringing across the board, while you should also observe slightly more dark detail being recovered. You may even see some recoverable stars appearing ever so slightly more pin-point.
At worst, when given a challenging dataset, deringing performance should be a wash between the two versions, but 550+ will always recover more dark detail.
- ST_1_9_536vs1.jpg (74.2 KiB) Viewed 504 times
The 1.9.550+ image (bottom) shows the sort of improvement you should generally be seeing; e.g. the image exhibits less ringing around a number of stars, while at the same time showing more structural detail/clarity in the darker areas in the nebulosity/filaments.
It is important, however, not to confuse pin-point with acuity; with 536 being more prone to ringing and "shadows", the local darkening around some stars would provide increased contrast (exactly like old school Unsharp mask). Acuity is obviously not the goal or purpose of deconvolution.
I'd be very interested to get my hands on a dataset where you feel there is a substantial regression.
So far I've just tested with one widefield galaxy dataset of limited integration (can provide if helpful), but tonight will try to start experimenting and comparing with more and different datasets.
If you'd like to share that dataset, it may indeed prove useful!