Err, this is far from neural networks. I once tried to combine it with some thing that can do the optimization faster, so that I could make it a tree structure (as to prevent feedback form amplifying iteratively, kind of like how you need to be careful that your GAN doesn't start doing dog pictures instead of cat pictures), and run the images that compromise a tree against the best guess of it's sibling tree, as well as vice versa, before using these new estimates of the distortion to get a new best guess from all images contained by the parent of these sibling trees. Repeat until you reach the top. You can use more than one sibling to handle non-power-of-two framecounts. There is partial software on my Github, in case someone is interested I can be reached, and while working for free isn't really in my interested (other things make more fun/seem more promising), I'd be happy to start working on it again if there was a reason.
Regarding your paper, I have to remind you that Michael got nice results from upsampling the images before running his software. I actually planned on using the texture units for this, to save on bandwidth/address calculation overhead in the pending partial rewrite of my software.
The GAN there also uses just a single frame, whereas this uses the properties of the distribution of the distortions when seen in the frequency domain to figure out how the distortions are most likely, and then combines the SNR from the many frames to a single image. There is research using a method very similar to Michael's with a GPU, GTX 580 or so iirc, which does >15 fps @720p in real time, with less than 2 frames latency and no more than 1 frame necessary latency if you run the GPU work queues rather empty (risking underutilisation if you don't get CPU time fast enough again). Combine with e.g. a nice Volta DGX, and something like a 400mm Schmidt camera including a field flattening lens and a CMOSIS CMV12000 (like, take the sensor out of an AXIOM beta camera, shrink the board around it to the smallest you can get, and stick it with a lens on top facing a 20 cm spherical mirror, with a corrective plate ~80cm from the mirror. This is about ~1000$ optics, 2500$ image hardware (including that necessary to get the full stream at >100 fps into the DGX), and whatever rent you pay for the DGX. Distortion free 10x slow motion with a pixel size of 14mm at 1km distance.
If you'd want to sell such a thing to non-military...
[0]: KIM, Dongmin; SRA, Suvrit; DHILLON, Inderjit S. A non-monotonic method for large-scale non-negative least squares. Optimization Methods and Software, 2013, 28. Jg., Nr. 5, S. 1012-1039.
https://pdfs.semanticscholar.org/622c/84cfba9781ad846105f28d...