Impact of AI-based advances in processing? Soul nebula

[timh]

Probably many folk have been looking at evaluating the recent RC-Astro NoiseXterminator and BlurXterminator (TM) tools for image processing - because of their apparent novelty of approach and the advance in image quality and simplicity that they may therefore bring?

As I understand it both tools were built using a training set of 'correct' HST etc. astro images to inform a neural net 'merit function' type approach to more accurately discriminate noise from signal(NoiseXt) or- in the case of BlurXt - to better estimate the 'true' target PSF function for LR deconvolution to minimise towards.

So if anyone finds it useful here is my partial evaluation based upon using some past data.

1) NoiseXterminator appears to do a very impressive job. It apparently increases SNR. But of course you can ask how would one ever know if it hadn't behaved correctly? So it is not transparent to the user exactly what it has done or exactly how - but that of course is in the nature of using a neural net. For myself, I trust the approach and like the improvement it brings -- but others may not trust it and it will be interesting to see how the tool evolves through further neural net updates.

2) BlurXterminator I see as a more incremental advance on Lucy Richardson deconvolution which is anyway included in PixInsight and is a tool that I have used quite often. Therefore I looked at a number of past images and compared the BlurXt results with the my previous best efforts at PI deconvolution.

To cut a long story short, BlurXt was far simpler to use and generally provided better results than did PI deconvolution (maybe 30%?) on most objects. It also has features that deal well with stars and there is no need to mask off the noisier areas -everything is dealt with more smoothly. However it of course can't do miracles (e.g. don't look for 2 arcsec resolution in an image sampled at 2 arcsec etc or great details from poor SNR etc). BlurXt is most appropriately applied to linear images at the earliest stage of processing. While it was superb on Type II nebula and more extended objects such as the ring nebula it is not yet able to work with very small bright objects such as the catseye and blinking planetary nebula (I sent these data to Russell Croman - and he agreed that these were out of scope currently) - so PI deconvolution not completely superceded yet.

Anyway to cut to the chase -- here is an example of what the software can do. Exactly the same Sharpcap acquired dataset as before on the soul nebula as described before -- viewtopic.php?t=6076.

The full old and new images are both here for comparison https://www.astrobin.com/bpa8vm/H/

For my own part, I think that the improvement is really quite marked. The new software is a gamechanger in one sense at least. Previously I would have looked at my IC1848 image and have been tempted to collect some more data to improve SNR, colour depth etc just that little bit. Now I am not - the software means that the image I already have is probably 'maxed out' for my particular telescope operating under Bortle 6 skies and so it is time to move on to a different field or object etc.

Tim


Herewith comparing close-up detail of the data processed as originally and as now using BlurXt and NoiseXt in the workstream

[MarMax]

I've been following the CN post on BxT and it sounds like a "must have" tool. For me, I'm not a PI user nor do I plan on purchasing and learning it. I have too many hours invested in Photoshop to change horses, plus I'm only doing EAA. It's great that NxT and SxT are Photoshop plugins and it's sad that BxT is not planned to be developed as a Photoshop plugin.

Thanks for your review and I'll just keep hoping for that Blue Moon event and RC developing a BxT Photoshop plugin.