Try increasing gamma if dark sections aren't distinguished

Try increasing gamma if dark sections aren't distinguished

Thursday, January 11, 2024

AI meets Jupiter

 Warning: boring details regarding planetary image processing ahead.  

Planetary imaging is best done these days with a web cam capturing short video segments with as many frames as possible.  software then combines the sharpest frames, discarding frames distorted by atmospheric turbulence.  the "stacked" image is that further sharpened.  too much sharpening produces unsightly artifacts enhancing noise inherent in the imaging process, rather than fine detail.  too little sharpening leaves a blurry image.  recently "AI" techniques have been used to sharpen images or "denoise" sharpened images.  the results have been very impressive, but controversial.  

Here's an image of jupiter with very conservative sharpening.  the image is nice and smooth, but blurry:

Jupiter 12/16/2023 05:20 UTC

next up is an image with aggressive sharpening applied.  fine details are more evident, but there are clearly sharpening artifacts, sometimes referred to as "noise":  


next a conventional "denoise" has been applied, removing the artifact, blurring the image slightly, but preserving some of the sharpening for a smoother sharpened image:


here's what topaz denoise AI did to the sharpened image:


the image is far more aesthetically pleasing than the others, both silky smooth and razor sharp.  this is what has generated the controversy.  there are 2 main objections:

1. the image is too smooth.  
-personally i think this is silly. hubble images of jupter are smooth and sharp at this scale, why not mine?  

2. the AI generates structures that are not real.  
-now this is a problem.  a number of advanced imagers have demonstrated AI processed images with structures that look real, but aren't actually there.  pretty damning criticism, thus all imagers submitting images for scientific work are avoiding AI processing.  i think you can make the case that the false AI structures are the result of pushing the AI processing too far, but now that the objection has been raised, AI processing is widely condemned in the planetary imaging circles.  

Image Details:

i'm submitting images to ALPO so now include image detail in the frame for the submission.  


cloudy nights thread on the topic--mostly civil, only one post was banned (but you can still see it's residue)

2 comments:

  1. Always interesting stuff. How does the software decide on what image is sharper than another. Also, how do the images grt aligned so that it's stacking the same location. I can imagine if that stacking images that were slightly offset from each ither would introduce some blurring

    ReplyDelete
  2. just an end user, so not sure, it sorts and ranks each frame. you can scroll through and actually see that the lower ranked images are much more blurry. the stacks are usually obviously sharper than the best single frame.
    https://www.astronomie.be/registax/
    https://en.wikipedia.org/wiki/Registax
    https://en.wikipedia.org/wiki/Shift-and-add

    ReplyDelete