I'm trying to use Topaz Video Enhance AI to upscale red and blue channels in order to create a better RGB video from RAW.
From one RAW file you can create R and B channels with half resolution, and then upscale them with AI into full resolution.
As for the G channel I just use averaging. It works quite nicely (no aliasing on EOS M and almost no color noise) however I ran into an issue: although I can split RAW into R, B and G fast enough, I can't do averaging for missing G pixels as fast enough with my tools. There must be some kind of RAW editor to do that, but I found mostly python and C code which is less desirable. Any suggestions?
It could be a great idea to directly train neural networks to upscale from raw data. It could lead to better performance than normal upscaling
In some other thread i posted similar idea - to train neural network with pairs of aliased and normal files (better on the raw level) to eliminate moire and aliasing, but seem no one interested in it
For fast training sample creation it is need to mimic/emulate the line skipping algo thats in the camera, and train with that pairs of raw data to reconstruct original (or create some new) information like Topaz Gigapixel do
There's loads of papers about neural network demosaicing if you search on google scholar.
It's a waste of time unless you're some kind of scientist imo.
Quote from: SoulState on May 11, 2023, 06:04:51 PM
In some other thread i posted similar idea - to train neural network with pairs of aliased and normal files (better on the raw level) to eliminate moire and aliasing, but seem no one interested in it
It's not that nobody is interested, it's that it's expensive and difficult.
I'll learn how to do it if you give me several million good image pairs, and donate the $100k of specialised hardware required.