This repository is the official PyTorch implementation of "Efficient Dual Attention Transformer for Image Super-Resolution" (ACM SAC 2024).

[paper] [visual results] [pretrained models]

Abstract : Research based on computationally efficient local-window selfattention has been actively advancing in the field of image superresolution (SR), leading to significant performance improvements. However, in most recent studies, local-window self-attention focuses only on spatial dimension, without sufficient consideration of the channel dimension. Additionally, extracting global information while maintaining the efficiency of local-window self-attention, still remains a challenging task in image SR. To resolve these problems, we propose a novel efficient dual attention transformer (EDAT). Our EDAT presents a dual attention block (DAB) that empowers the exploration of interdependencies not just among features residing at diverse spatial locations but also among distinct channels. Moreover, we propose a global attention block (GAB) to achieve efficient global feature extraction by reducing the spatial size of the keys and values. Our extensive experiments demonstrate that our DAB and GAB complement each other, exhibiting a synergistic effect. Furthermore, based on the two attention blocks, DAB and GAB, our EDAT achieves state-of-the-art results on five benchmark datasets.

SR (x4)	HQ	SwinIR	CAT-A	EDAT (Ours)

• python 3.8
• pyTorch >= 1.8.0
• NVIDIA GPU + CUDA

git clone https://github.com/soobin419/EDAT.git
cd EDAT
pip install -r requirements.txt
python setup.py develop

Used training and testing sets can be downloaded as follows:

Download training and testing datasets and put them into the corresponding folders of datasets/. See datasets for the detail of the directory structure.

• Download training (DF2K, already processed) and testing (Set5, Set14, BSD100, Urban100, Manga109, already processed) datasets, place them in datasets/.

• Run the following scripts. The training configuration is in options/train/.

  # EDAT, input=64x64, 4 GPUs, batch size=8 per GPU
  python -m torch.distributed.launch --nproc_per_node=4 --master_port=2414 basicsr/train.py -opt options/train/train_EDAT_SRx2_from_scratch.yml --launcher pytorch
  python -m torch.distributed.launch --nproc_per_node=4 --master_port=2414 basicsr/train.py -opt options/train/train_EDAT_SRx3_finetune_from_SRX2.yml --launcher pytorch
  python -m torch.distributed.launch --nproc_per_node=4 --master_port=2414 basicsr/train.py -opt options/train/train_EDAT_SRx4_finetune_from_SRX2.yml --launcher pytorch

• The training experiment is in experiments/.

• Download the pre-trained models and place them in experiments/pretrained_models/.

  We provide pre-trained models for image SR: EDAT (x2, x3, x4).

• Download testing (Set5, Set14, BSD100, Urban100, Manga109) datasets, place them in datasets/.

• Run the following scripts. The testing configuration is in options/test/ (e.g., test_SR_X2.yml).

  # EDAT, reproduces results in Table 2 of the main paper
  python basicsr/test.py -opt options/test/test_SR_X2.yml
  python basicsr/test.py -opt options/test/test_SR_X3.yml
  python basicsr/test.py -opt options/test/test_SR_X4.yml

• The output is in results/.

• Download the pre-trained models and place them in experiments/pretrained_models/.

  We provide pre-trained models for image SR: EDAT (x2, x3, x4).

• Put your dataset (single LR images) in datasets/single. Some test images are in this folder.

• Run the following scripts. The testing configuration is in options/test/ (e.g., test_EDAT_single_x2.yml).

  Note 1: The default model is DWT. You can use other models like DWT-S by modifying the YML.

  # Test on your dataset
  python basicsr/test.py -opt options/test/test_EDAT_single_x2.yml
  python basicsr/test.py -opt options/test/test_EDAT_single_x3.yml
  python basicsr/test.py -opt options/test/test_EDAT_single_x4.yml

• The output is in results/.

We achieved state-of-the-art performance. Detailed results can be found in the paper. All visual results of EDAT can be downloaded here.

This code is built on BasicSR.