Home

Awesome

MoESR: Blind Super-Resolution using Kernel-Aware Mixture of Experts

This repository is the official implementation of "MoESR: Blind Super-Resolution using Kernel-Aware Mixture of Experts".

Requirements

To install requirements:

conda env create -f environment.yml
conda activate MoESR_env

Datasets and pretrained models

You can download all datasets (DIV2KRK, Flickr2KRK and Urban100RK) and pretrained models from the following link: https://drive.google.com/drive/folders/1v7Lthkp-nLwdXGGkqolBgl5H88oNDH6l?usp=sharing

Test on synthetic datasets

For example to evaluate on DIV2KRK dataset:

cd codes
python main.py --in_dir ../datasets/DIV2KRK/lr_x2 --out_dir ../results/DIV2KRK/x2 --gt_dir ../datasets/DIV2KRK/gt --scale 2
python main.py --in_dir ../datasets/DIV2KRK/lr_x4 --out_dir ../results/DIV2KRK/x4 --gt_dir ../datasets/DIV2KRK/gt --scale 4

Test on real images

To evaluate on real-world images:

cd codes
python main.py --in_dir 'path to the LR input images' --out_dir 'path to save results' --scale 2 --real
python main.py --in_dir 'path to the LR input images' --out_dir 'path to save results' --scale 4 --real

Results

Our model achieves the following performance values (PSNR / SSIM) on DIV2KRK, Flickr2KRK and Urban100RK datasets:

Model nameScaleDIV2KRKFlickr2KRKUrban100RK
MoESRx232.69 / 0.905432.95 / 0.905627.29 / 0.8448
MoESRx428.48 / 0.780528.57 / 0.779523.62 / 0.6766

Acknowledgement

The code is built on DualSR and KernelGAN. We thank the authors for sharing the codes.