Image Super Resolution

Hi, is there any Julia project focusing on Image Super Resolution? There are lots of these libraries for e.g. Python or Mathematica, but found not one for Julia:

I’m assuming you’re referring to super resolution specifically in terms of neural networks.

What you post is likely outdated (ESRGAN from 2018, vs from 2020 I link to below, see Tencent link for impressive pictures, and the problems they solved, “[old] methods always fail in real-world image super-resolution”). If you want to reimplement something in Julia, please use state-of-the-art, such as I link to below (and you likely do not have to, can just call Python code?).

First what I see is available in Julia (I didn’t look closely, it’s probably not based on neural-network):

The slides here:
https://web.eecs.umich.edu/~fessler/papers/files/talk/19/ima-julia/
Slide 15: Super-resolution example

using MIRT: jim, ellipse_im


See at: https://web.eecs.umich.edu/~fessler/papers/files/talk/19/ima-julia/

Shreyas Kowshik: Addition Of Baseline Models To Flux.jl

The current state of the art algorithms in terms of reinforcement learning and generative models are not yet available in Flux. This project aims to add the following models with explicit documentation in the form of blog and code :

  • […]
  • Super Resolution GAN

Interesting links, first what I knew about, then something I dug up now:

In addition, our method is the winner of NTIRE 2020 Challenge on both tracks of Real-World Super-Resolution, which significantly outperforms other competitors by large margins.

This one also scored high:

And something specific to faces might be useful:

Didn’t know about this, it’s pretty neat:

2 Likes

Some additional links I had.

See there more amazing upscaling pictures, and source code (for Tencent link above), and executables for Linux Windows, macOS (that you could use independently, or from Julia or the underlying C++ library, even on Android and iOS too):

Upscale 4x with ImageMagick Lanczo4 Filter
[vs]
Upscale 4x with realsr model=DF2K scale=4 tta=1
realsr-ncnn-vulkan.exe -i origin.jpg -o output.png -s 4 -x -m models-DF2K

“-resize 400%”, i.e. 4x may be standard terminology, or not, for 16x larger.

FYI: I also see the term “Extreme Super-Resolution” used, but I believe that’s the same magnification as above, as I do not really believe 16x16=256x magnification:

Abstract:

This paper reviews the NTIRE 2020 challenge on perceptual extreme super-resolution with focus on proposed solutions and results. The challenge task was to super-resolve an input image with a magnification factor Ă—16 based on a set of prior examples of low and corresponding high resolution images.

See also here, where you can compare 7 methods across several pictures:

Blind Super-Resolution Methods comparison

http://www.wisdom.weizmann.ac.il/~vision/kernelgan/resources/more_examples.html