|
I think AI in processing astrophotography is likely to be the equivalent for some time of CHAT GPT.
By that I mean they have already shown that CHAT GPT is quite capable of producing an essay to argue a point for instance, that reads well, hangs together grammatically, sounds authoritative but in terms of the subject matter is completely wrong. I was listening to talk on the radio about it and they were talking about in education settings, possibly having to go to oral exams when they have doubts, if a student can go feed a few text prompts at an AI program and get back a well-structured essay that is contextually right (And surely eventually they will) then how do you work out if the student actually understands the subject matter except for giving them a grilling face to face.
The AP equivalent is some of the AI tools which simply features where none exist, they look great, look like details but they come from an unknowable association in "Black box" software. Make sure it has the right training set or the Vela SNR suddenly looks like the "Wall of noses" in Jenolan caves.
The other side of the coin is that for AP it might well eventually be able to take whatever rubbish data you throw at it, identify the object and massage your data until the "Output" looks like the reference image.
Neither option appeals to me in the least. Star removal by AI like StarXterminator is one thing, and I am fairly certain it is inventing detail as it "Removes" the stars (If the star has saturated the pixels in an area for instance, no amount of extrapolation is anything but a guess at what would be there if the star was not) but my use of that type of tool is generally to remove stars from narrowband images before using them as "Lighten" layers to add the NB to LRGB images, the stars in the LRGB are going to push through and obscure the invented data anyway. I also like NoiseXterminator, but that has to have limits too before it is interpolating away detail that is actually there, having mis-identified it as noise.
TL;DR:
I reckon beyond a certain point (Which we are pretty close to IMO) AI tools in AP will not improve the breed. It is not inconceivable to have an AI set up that can identify objects by plate solving and corral whatever rubbish you throw at it into a good image. But is getting out your pencil and drawing a star field on paper for it to solve, then letting it massage your coffee stains into an image of M42 astrophotography?
|