Lightweight Single-Image Super-Resolution Network with Attentive Auxiliary Feature Learning

An official implementation about our paper: Lightweight Single-Image Super-Resolution Network with Attentive Auxiliary Feature Learning(ACCV2020)

Paper(Arxiv)

Project Page

Dependecies

python = 3.6
pytorch = 1.1.0 (higher is OK)
torchvision = 0.3.0
cudatoolkit = 9.0
tqdm = 4.48.0
easydict = 1.9
scikit-image = 0.17.2
opencv-contrib-python = 4.3.0.36
pyyaml = 5.3.1
pillow = 6.2.2
scipy = 1.2.1

If you use Anaconda, you can run these commands to build a env that can run our code correctly:

#First step: build a env and enter it
conda create -n AAF python=3.6
conda activate AAF

#Second step: install pytorch and torchvision via conda
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch

#Final step: install other dependecies via pip
pip install -r requirements.txt

Now, you have build a conda env for running code successfully!

Training

Firstly, please refer to DIV2K to get DIV2K which is a widely-used dataset for super-resolution. Put it in any folder, and modify the parameter dataset.dir_root in yaml(e.g. yaml/train_sd_x4.yaml) as this folder. Our method will scan files according to dataset.dir_root/dataset.dir_lr.

Perform below script to train the model from scratch.

python train.py --config AAF_SD_x4  #AAF-SD for x4 scale

Here, AAF_SD_x4 can be change as:

AAF_S_x4    #AAF-S for x4 scale
AAF_SD_x4   #AAF-SD for x4 scale
AAF_M_x4    #AAF-M for x4 scale
AAF_L_x4    #AAF-L for x4 scale
AAF_S_x3    #AAF-S for x3 scale
AAF_SD_x3   #AAF-SD for x3 scale
AAF_M_x3    #AAF-M for x3 scale
AAF_L_x3    #AAF-L for x3 scale
AAF_S_x2*   #AAF-S for x2 scale
AAF_SD_x2   #AAF-SD for x2 scale
AAF_M_x2*   #AAF-M for x2 scale
AAF_L_x2*   #AAF-L for x2 scale

Testing

Please put the testing data (such like DIV2K_test_LR_bicubic/X4) that you want to check results in ./test_img, now you will have:

|- ./test_img/
|--- DIV2K_test_LR_bicubic/
|------ X4/
|--------- 0901x4.png
|--------- 0902x4.png
|--------- ...

We provide the pretrained model of AAF in ./checkpoint, so you can test images directly. Run this commands to obatain the results. They can be seen in ./results/AAF_SD_x4:

python test.py --config AAF_SD_x4

Others

Paramaters

Parameters used in our code can be seen in ./yaml/*.yaml.

*About pretrained model

Unfortunately, we encounter a problem with our hard disk, which cause a loss of data including the pretrained model (x2 settings except AAF-SD). We will release the other x2 checkpoints later.

Citation

If you think the paper is helpful for your research, please cite:

@inproceedings{wang2020Lightweight,
        author = {Xuehui Wang, Qing Wang, Yuzhi Zhao, Junchi Yan, Lei Fan, and Long Chen.},
        title = {Lightweight Single-Image Super-Resolution Network with Attentive Auxiliary Feature Learning},
        journal = {ACCV},
        year = {2020}
}

If you have any question, please contact wangxh228@mail2.sysu.edu.cn

tamwaiban/A2F-SR