README.md

Closed-loop Matters: Dual Regression Networks for Single Image Super-Resolution [arXiv]

Pytorch implementation for "Closed-loop Matters: Dual Regression Networks for Single Image Super-Resolution".

Important Update

• We released the code and models of our Dual Regression Compression (DRC) towards lightweight image SR.
• Our DRC paper is already available on arXiv.

Dependencies

Python>=3.7, PyTorch>=1.1, numpy, skimage, imageio, matplotlib, tqdm

Quickstart (Model Testing)

Results of our pretrained models:

Model	Scale	#Params (M)	PSNR on Set5 (dB)
DRN-S	4	4.8	32.68
	8	5.4	27.41
DRN-L	4	9.8	32.74
	8	10.0	27.43

You can evaluate our models on several widely used benchmark datasets, including Set5, Set14, B100, Urban100, Manga109. Note that using an old PyTorch version (earlier than 1.1) would yield wrong results.

Please organize the benchmark datasets using the following hierarchy.

- srdata
    - benchmark
         - Set5
            - LR_bicubic
                - X4
                    - babyx4.png

You can use the following script to obtain the testing results:

python main.py --data_dir $DATA_DIR$ \
--save $SAVE_DIR$ --data_test $DATA_TEST$ \
--scale $SCALE$ --model $MODEL$ \
--pre_train $PRETRAINED_MODEL$ \
--test_only --save_results

• DATA_DIR: path to save data
• SAVE_DIR: path to save experiment results
• DATA_TEST: the data to be tested, such as Set5, Set14, B100, Urban100, and Manga109
• SCALE: super resolution scale, such as 4 and 8
• MODEL: model type, such as DRN-S and DRN-L
• PRETRAINED_MODEL: path of the pretrained model

For example, you can use the following command to test our DRN-S model for 4x SR.

python main.py --data_dir ~/srdata \
--save ../experiments --data_test Set5 \
--scale 4 --model DRN-S \
--pre_train ../pretrained_models/DRNS4x.pt \
--test_only --save_results

You will obtain the output like this.

If you want to load the pretrained dual model, you can add the following option into the command.

--pre_train_dual ../pretrained_models/DRNS4x_dual_model.pt

Training Method

We use DF2K dataset (the combination of DIV2K and Flickr2K datasets) to train DRN-S and DRN-L.

python main.py --data_dir $DATA_DIR$ \
--scale $SCALE$ --model $MODEL$ \
--save $SAVE_DIR$

• DATA_DIR: path to save data
• SCALE: super resolution scale, such as 4 and 8
• MODEL: model type, such as DRN-S and DRN-L
• SAVE_DIR: path to save experiment results

For example, you can use the following command to train the DRN-S model for 4x SR.

python main.py --data_dir ~/srdata \
--scale 4 --model DRN-S \
--save ../experiments 

Citation

If you use any part of this code in your research, please cite our paper:

@inproceedings{guo2020closed,
  title={Closed-loop Matters: Dual Regression Networks for Single Image Super-Resolution},
  author={Guo, Yong and Chen, Jian and Wang, Jingdong and Chen, Qi and Cao, Jiezhang and Deng, Zeshuai and Xu, Yanwu and Tan, Mingkui},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2020}
}