4PP-EUSR

Four-pass perceptual super-resolution with enhanced upscaling

Introduction

This repository contains a TensorFlow-based implementation of 4PP-EUSR ("Deep learning-based image super-resolution considering quantitative and perceptual quality"), which considers both the quantitative (e.g., PSNR) and perceptual quality (e.g., NIQE) of the upscaled images. Our method won the 2nd place and got the highest human opinion score for Region 2 in the 2018 PIRM Challenge on Perceptual Image Super-resolution at ECCV 2018.

※ The perceptual index is calculated by "0.5 * ((10 - Ma) + NIQE)", which is used in the PIRM Challenge. Lower is better.

Followings are the performance comparison evaluated on the BSD100 dataset.

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

J.-H. Choi, J.-H. Kim, M. Cheon, J.-S. Lee: Deep learning-based image super-resolution considering quantitative and perceptual quality. arXiv:1809.04789 (2018)
J.-H. Kim, J.-S. Lee: Deep residual network with enhanced upscaling module for super-resolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pp. 913-921 (2018)

Generating upscaled images from the trained models can be done by test/test.py. Following are the brief instructions.

Download and copy the trained model available in Downloads section to the test/ folder.
Place the low-resolution images (PNG only) to the test/LR/ folder.
Run python test.py --model_name [model file name]. For example, if you downloaded the PIRM Challenge version of our pre-trained model, run python test.py --model_name 4pp_eusr_pirm.pb.
The upscaled images will be available on the test/SR/ folder.

Please run python test.py --help for more information.

Implementing the training code is in progress. Please refer to README.md on the train/ folder for more information.

Pre-trained models:

Results (Set5, Set14, BSD100, PIRM):