search

jixiaozhong / RealSR

Real-World Super-Resolution via Kernel Estimation and Noise Injection
Apache License 2.0
738 stars 102 forks source link

About the 'Kernel Estimation' and 'Noise collection ' #13

Open babyjia opened 4 years ago

babyjia commented 4 years ago

I have not found 'Kernel Estimation' and 'Noise collection ' in data codes, is it not included in release codes so far?

YunjinChen commented 4 years ago

同问+1

fuuuyuuu commented 4 years ago

"Noise collection" is hard to understand through this paper ... but "Kernel Estimation" can be found in paper "Blind Super-Resolution Kernel Estimation using an Internal-GAN"

ALLinLLM commented 3 years ago

@fuuuyuuu I try 3 kinds of Noise collection to create "noise pool", they are:

  1. doble slide window in Image Blind Denoising With Generative Adversarial Network Based Noise Modeling
  2. NoiseGAN in Kernel Modeling Super-Resolution on Real Low-Resolution Images
  3. ||mean(pixel) - std(pixel)|| in Noise2Blur: Online Noise Extraction and Denoising

However, @jixiaozhong I failed to achieve the pretrained realSR model's perceptual quality level 😭 really want to learn the way you train the model

jixiaozhong commented 3 years ago

The noise cut variance is set as 20, and the patch size is set as 256. Convert RGB to Gray and calculate the variance. We are open source code through the company's channels, and the training code should be released soon (maybe a week or two).