FDIWN-Pytorch

This repository is an official PyTorch implementation of our paper "Feature Distillation Interaction Weighting Network for Lightweight Image Super-Resolution". (AAAI 2022)

Paper can be download from FDIWN.

All test datasets (Preprocessed HR images) can be downloaded from here.

All original test datasets (HR images) can be downloaded from here.

Prerequisites:

1. Python 3.6
2. PyTorch >= 0.4.0
3. numpy
4. skimage
5. imageio
6. matplotlib
7. tqdm

Dataset

We used DIV2K dataset to train our model. Please download it from here.

Extract the file and put it into the Train/dataset.

Only DIV2K is used as the training dataset, and Flickr2K is not used as the training dataset !!!

Results

All our SR images can be downloaded from here.[百度网盘][提取码:0824]

All our Supplementary materials can be downloaded from here.[百度网盘][提取码:9168]

All pretrained model can be found in AAAI2022_FDIWN_premodel.

The following PSNR/SSIMs are evaluated on Matlab R2017a and the code can be referred to [Evaluate_PSNR_SSIM.m] (https://github.com/24wenjie-li/FDIWN/blob/main/FDIWN_TestCode/Evaluate_PSNR_SSIM.m).

Training

Don't use --ext sep argument on your first running.

You can skip the decoding part and use saved binaries with --ext sep argument in second time.

  cd Train/
  # FDIWN x2  LR: 48 * 48  HR: 96 * 96
  python main.py --model FDIWNx2 --save FDIWNx2 --scale 2 --n_feats 64  --reset --chop --save_results --patch_size 96 --ext sep
  
  # FDIWN x3  LR: 48 * 48  HR: 144 * 144
  python main.py --model FDIWNx3 --save FDIWNx3 --scale 3 --n_feats 64  --reset --chop --save_results --patch_size 144 --ext sep
  
  # FDIWN x4  LR: 48 * 48  HR: 192 * 192
  python main.py --model FDIWNx4 --save FDIWNx4 --scale 4 --n_feats 64  --reset --chop --save_results --patch_size 192 --ext sep

Testing

Using pre-trained model for training, all test datasets must be pretreatment by ''FDIWN_TestCode/Prepare_TestData_HR_LR.m" and all pre-trained model should be put into "FDIWN_TestCode/model/".

#FDIWN x2
python main.py --data_test MyImage --scale 2 --model FDIWNx2 --n_feats 64 --pre_train /home/ggw/wenjieli/RCAN/RCAN_TestCode/model/model_best.pt --test_only --save_results --chop --save FDIWNx2 --testpath ../LR/LRBI --testset Set5

#FDIWN x3
python main.py --data_test MyImage --scale 3 --model FDIWNx3 --n_feats 64 --pre_train /home/ggw/wenjieli/RCAN/RCAN_TestCode/model/model_best.pt --test_only --save_results --chop --save FDIWNx3 --testpath ../LR/LRBI --testset Set5

#FDIWN x4
python main.py --data_test MyImage --scale 4 --model FDIWNx4 --n_feats 64 --pre_train /home/ggw/wenjieli/RCAN/RCAN_TestCode/model/model_best.pt --test_only --save_results --chop --save FDIWNx4 --testpath ../LR/LRBI --testset Set5

Performance

Our FDIWN is trained on RGB, but as in previous work, we only reported PSNR/SSIM on the Y channel.

We use the file ''...FDIWN_TestCode/Evaluate_PSNR_SSIM'' for test.

Model	Scale	Params	Multi-adds	Set5	Set14	B100	Urban100	Manga109
FDIWN-M	x2	433K	73.6G	38.03/0.9606	33.60/0.9179	32.17/0.8995	32.19/0.9284	null/null
FDIWN	x2	629K	112.0G	38.07/0.9608	33.75/0.9201	32.23/0.9003	32.40/0.9305	38.85/0.9774
FDIWN-M	x3	446K	35.9G	34.46/0.9274	30.35/0.8423	29.10/0.8051	28.16/0.8528	null/null
FDIWN	x3	645K	51.5G	34.52/0.9281	30.42/0.8438	29.14/0.8065	28.36/0.8567	33.77/0.9456
FDIWN-M	x4	454K	19.6G	32.17/0.8941	28.55/0.7806	27.58/0.7364	26.02/0.7844	null/null
FDIWN	x4	664K	28.4G	32.23/0.8955	28.66/0.7829	27.62/0.7380	26.28/0.7919	30.63/0.9098

Model complexity

FDIWN gains a better trade-off between model size, performance, inference speed, and multi-adds.

Citation

If you find FDIWN useful in your research, please consider citing:

@article{gao2021feature,
  title={Feature Distillation Interaction Weighting Network for Lightweight Image Super-Resolution},
  author={Gao, Guangwei and Li, Wenjie and Li, Juncheng and Wu, Fei and Lu, Huimin and Yu, Yi},
  journal={arXiv preprint arXiv:2112.08655},
  year={2021}
}