Neural Enhancer

[Tippon]

@edps Did you ever get Neural Enhancer working properly? I’ve set it up on Ubuntu Studio 16.04, but I get a syntax error when I try to run it.

I’ve got a .pdf of the MM thread on it that I’ve been working from, but I think I must have missed a step somewhere.

 

EDIT: I’ve got it working now by installing Docker, but only Zoom makes any noticeable difference. The Repair option very slightly brightens small areas of the photo, Deblur ruins the photo, and Denoise claims that support files are missing  :wacko:

[Ed P]

I had similar experiences to you. Nice idea, shame about the output!

I could not work out the basis used for their training pictures. I think the flaw was probably in the way they paired low and high resolution pics of the same scene.

[Tippon]

I did have a look at downloading the training data and retraining it myself, but then realised that they’re currently using over a million images, and their example shows the download maxing out a 200Mb connection for several days and still going!  :unsure:

I’ve got about 700 photos from my time working in a nightclub that I want to enlarge and enhance, so I might kill a bit of time by cloning them to up the numbers to a few thousand, and see how that affects the training.

[Ed P]

The problem I have with it is that  the typical way of training a neural net is that you have (say) a fuzzy out of focus pic of a scene and a good well focused pic of the same screen, you then convolve the fuzzy pic using a zillion different convolution matrices and rate the best matrix  fit. Rinse & repeat for a zillion different pics. That trained convolution matrix is then the one that is used in the production run. I could not identify that sort of process going on within the training routine. One day I will have to reread that part of the program to see what it is happening.

[Author]

Posts