Home

Awesome

EUSR-PCL

GAN-based image super-resolution using perceptual content losses (tensorflow implementation)

Introduction

This repository include a implementation of GAN-based single image super-resolution using perceptual content losses (PCL), which considers both distortion- and perception-based quality of super-resolved images. In the PIRM Challenge on Perceptual Super Resolution at ECCV 2018, Our team (Yonsei-MCML) won the 2nd place for Region 1. (Our team also won the 2nd place for Region 2 based on 4PP-EUSR model.)

Please cite following papers when you use the code, pre-trained models, or results:

Performance of the method

The perceptual index is calculated by two no-reference quality measurements, Ma and NIQE. Lower score means better perceptual quality. The detail of this index is explained in the PIRM Challenge.

Final ranking of our method (Yonsei-MCML) (please check the details in PIRM website)

Final ranking

Usage of testing code

The instructions for the usage of testing code is below. Generating super-resolved images from the pre-trained models can be done by <test/test.py>. Now, we only support x4 super-resolution for the challenge.

  1. Download and copy the trained model available in Downloads section to the <test/> folder.
  2. Place the low-resolution images (PNG files) to the <test/LR/> folder.
  3. Run <python test.py>
  4. The super-resolved images will be available on the <test/LR/> folder.

Usage of training code

The training code will be upaded soon.

Downloads

Pre-trained models (for the PIRM Challenge)