BlurXTerminator - yes or no?

Recently, a new tool for PixInsight was released - Russel Coman's BlurXTerminator. Another AI based tool for hobby astronomers, this time for deconvolution - resulting in sharp stars and amazing details in galaxies and nebulae.

(See below some examples for reprocessing some of my images, i.e. the Rosette Nebula, Jellyfish Nebula and Elephant's trunk, using the workflow described here. BlurXTerminator was applied to the basic Ha and OIII images after stacking.)

This tool provides an AI solution for one of the most cumbersome steps in astronomy image processing.

We now have AI tools for

• star masking and removal (e.g. Starnet++, StarXterminator)
• noise removal (e.g. Topaz DeNoise AI, NoiseXterminator)
• and deconvolution.

Each of these tasks would typically require many manual steps, detailed knowledge of the tools and lots of experience for good results.

I do not want to contribute to the technical discussions about these tools, the background can be found on the vendors' web pages and in hundreds of forum threads in astronomy boards - but with the upcoming discussions about usage of AI (especially ChatGPT and its relatives), I would like to add my two cents (besides that, the spell and grammar checker for this blog is AI based, too :-))

I am not concerned about a competition on publishing the best images. Astrophotography is just a hobby for most of us - and if someone creates an income by selling images and now even better ones, why not? There is still plenty of work to be done for good images. We still need to select the targets, choose our best gear, set up the equipment, think about filters, imaging time, sub frames, darks, flats, processing the data properly, and finally manage the post-processing for beautiful pictures. Same spot, same night, same target - and each of us would produce a different wonderful image.

I am more concerned about what gets lost. If we look at how we are taking our images and how this might be done in the future, I am struggling with a growing deficiency in engineering background.

Let's list a couple of things, of which each of us is making use more or less:

• We select our objects with planetary software. I often feel like I am using a navigation system in an unfamiliar city - I know how to get to my target, but still do not know where I am. In the 80s I was out in the backyard with my phosphorescent star map and learned how to identify stars, messier objects, and constellations by heart.
• We are using motor driven mounts for tracking. With my first telescope in the 80s, I head to learn how to move two axes manually in Alt-Az mode or how to blindly track stars in EQ mode as my camera (with a film!) was blocking the eyepiece.
• We are using digital astro cameras, cooled to -20 °C, with high speed USB. For my first digital images I had to deal with a DSLR camera - with batteries which did not last for more than 45 minutes in a freezy night, image focussing with the camera display, shutter vibration and manual sub-frame timing.
• We are using focusers with stepper motors, resulting in super sharp images with the ability e.g. for temperature controlled re-focussing. There must be a big market for second hand Bahtinov masks, now.
• We calibrate our mounts using plate solving and calculating the deviation from the coordinate settings. No more need to search for landmark stars and centering them with a cross-hair eyepiece.
• We track stars with a guiding camera and guiding software, we do not have to take care about PEC errors, sometimes not even about bad polar alignment.

For sure, this list is not complete. Each of us takes advantage of a few if not all the improvements listed above. But not all of us have learned how to deal without them.

The same is now slowly creeping into our image processing workflow. If we take a look at the PixInsight tools, most of us will admit, that we do not know how they work in detail, when we set the parameters. But at least, there is always an algorithm, a formula, some mathematics we could understand, if we are interested in the background.

I must admit, I am a lame duck, when it comes to deconvolution. I never found the best parameters - and here I confess, I purchased a BlurXTerminator license because this final missing step in my tool chain was annoying me. There is still lots to learn with the tool, depending on the workflow. I am using it in different branches in the processing chain, with and without star sharpening, and I am always looking for the PSF which does the best job on the details of the nebula.

BUT (and this is the point): I will never learn how to do deconvolution in the old-fashioned way. And this is my biggest concern - that with all the AI tools around us, we forget the basics - but the basics were essential for the evolution of engineering. Without all the people developing these techniques and doing the engineering work, we would never have been able to develop or even train AI software. If we stop thinking about engineering, mathematics, algorithms, we are running at full-throttle into a dead-end. So if you use AI for image processing in astrophotography - take a step back from time to time and try to do it without. Even if you throw away the results - at least you still know how to navigate in a big city with a map in your hands.