Awesome

Our new work about blind image super-resolution has been accepted to IJCV. The paper is available at End-to-end Alternating Optimization for Real-World Blind Super Resolution. The codes are released at RealDAN.

This is an official implementation of Unfolding the Alternating Optimization for Blind Super Resolution and End-to-end Alternating Optimization for Blind Super Resolution

If this repo works for you, please cite our papers:

@article{luo2020unfolding,
  title={Unfolding the Alternating Optimization for Blind Super Resolution},
  author={Luo, Zhengxiong and Huang, Yan and Li, Shang and Wang, Liang and Tan, Tieniu},
  journal={Advances in Neural Information Processing Systems (NeurIPS)},
  volume={33},
  year={2020}
}
@misc{luo2021endtoend,
      title={End-to-end Alternating Optimization for Blind Super Resolution}, 
      author={Zhengxiong Luo and Yan Huang and Shang Li and Liang Wang and Tieniu Tan},
      year={2021},
      eprint={2105.06878},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

This repo is build on the basis of [MMSR] and [IKC]

News

Add more pretrained weights and update the results of DANv1 !
Add pretrained weights and update the results of about [IKC]!
Add DANv2 !!!

Main Results

Results about Setting 1

Method	Scale	Set5	Set5	Set14	Set14	B100	B100	Urban100	Urban100	Mangan109	Manga109
		PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
IKC	x2	37.19	0.9526	32.94	0.9024	31.51	0.8790	29.85	0.8928	36.93	0.9667
DANv1	x2	37.34	0.9526	33.08	0.9041	31.76	0.8858	30.60	0.9060	37.23	0.9710
DANv2	x2	37.60	0.9544	33.44	0.9094	32.00	0.8904	31.43	0.9174	38.07	0.9734
IKC	x3	33.06	0.9146	29.38	0.8233	28.53	0.7899	27.43	0.8302	32.43	0.9316
DANv1	x3	34.04	0.9199	30.09	0.8287	28.94	0.7919	27.65	0.8352	33.16	0.9382
DANv2	x3	34.19	0.9209	30.20	0.8309	29.03	0.7948	27.83	0.8395	33.28	0.9400
IKC	x4	31.67	0.8829	28.31	0.7643	27.37	0.7192	25.33	0.7504	28.91	0.8782
DANv1	x4	31.89	0.8864	28.42	0.7687	27.51	0.7248	25.86	0.7721	30.50	0.9037
DANv2	x4	32.00	0.8885	28.50	0.7715	27.56	0.7277	25.94	0.7748	30.45	0.9037

Results about Setting 2 (DIV2KRK)

Method	x2	x2	x4	x4
	PSNR	SSIM	PSNR	SSIM
KernelGAN + ZSSR	30.36	0.8669	26.81	0.7316
DANv1	32.56	0.8997	27.55	0.7582
DANv2	32.58	0.9048	28.74	0.7893

Dependenices

python3
pytorch >= 1.5
NVIDIA GPU + CUDA
Python packages: pip3 install numpy opencv-python lmdb pyyaml

Pretrained Weights

Pretrained weights of DANv1 and IKC are available at BaiduYun(Password: cbjv) or GoogleDrive. Download the weights to checkpoints

.
|-- checkpoints
`-- |-- DANv1
    |   |-- ...
    |-- DANv2
    |   |-- ...
    `--IKC
        |-- ...

Dataset Preparation

We use DIV2K and Flickr2K as our training datasets.

For evaluation of Setting 1, we use five datasets, i.e., Set5, Set14, Urban100, BSD100 and Manga109.

We use DIV2KRK for evaluation of Setting 2.

To train a model on the full dataset(DIV2K+Flickr2K, totally 3450 images), download datasets from official websites. After download, run codes/scripts/generate_mod_blur_LR_bic.py to generate LRblur/LR/HR/Bicubic datasets paths. (You need to modify the file paths by yourself.)

python3 codes/scripts/generate_mod_blur_LR_bic.py

For efficient IO, run codes/scripts/create_lmdb.py to transform datasets to binary files. (You need to modify the file paths by yourself.)

python3 codes/scripts/create_lmdb.py

Train

For single GPU:

cd codes/config/DANv1
python3 train.py -opt=options/setting1/train_setting1_x4.yml

For distributed training

cd codes/config/DANv1
python3 -m torch.distributed.launch --nproc_per_node=8 --master_poer=4321 train.py -opt=options/setting1/train_setting1_x4.yml --launcher pytorch

Test on Synthetic Images

cd codes/config/DANv1
python3 test.py -opt=options/setting1/test_setting1_x4.yml

Test on Real Images

cd codes/config/DANv1
python3 inference.py -input_dir=/path/to/real/images/ -output_dir=/path/to/save/sr/results/