Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution
Karam Park, Jae Woong Soh, and Nam Ik Cho
Accepted for AAAI 2025
Paper and Supplementary Material
- Ubuntu 18.04
- PyTorch 2.2.2
- CUDA 9.0 & cuDNN 7.1
- Python 3.10.9
Dependencies:
- PyTorch>1.10
- OpenCV
- Matplotlib 3.3.4
- opencv-python
- pyyaml
- tqdm
- numpy
- torchvision
We would like to express our thanks to the authors of Omni-SR for generously releasing their code to the public. Our codes are built on Omni-SR. If you encounter any problems using the code, please refer to Omni-SR Issue Threads first.
Transformer-based Super-Resolution (SR) methods have demonstrated superior performance compared to convolutional neural network (CNN)-based SR approaches due to their capability to capture long-range dependencies. However, their high computational complexity necessitates the development of lightweight approaches for practical use. To address this challenge, we propose the Attention-Sharing Information Distillation (ASID) network, a lightweight SR network that integrates attention-sharing and an information distillation structure specifically designed for Transformer-based SR methods. We modify the information distillation scheme, originally designed for efficient CNN operations, to reduce the computational load of stacked self-attention layers, effectively addressing the efficiency bottleneck. Additionally, we introduce attention-sharing across blocks to further minimize the computational cost of self-attention operations. By combining these strategies, ASID achieves competitive performance with existing SR methods while requiring only around 300K parameters - significantly fewer than existing CNN-based and Transformer-based SR models. Furthermore, ASID outperforms state-of-the-art SR methods when the number of parameters is matched, demonstrating its efficiency and effectiveness.
ASID delivers competitive performance compared to existing lightweight SR methods while utilizing significantly fewer model parameters.
More results are available in ./SR_Results.
Requisites should be installed beforehand.
-
Make sure the location configuration is correct in
./env/env.json -
Evaluate models with the following cmd
python test.py -v "Model_Name" -t tetser_Matlab -s 0 --test_dataset_name [Dataset]
[Model]: ASID_XN_DIV2K, ASIDd8_XN_DIV2K (N=2,3,4)
[Dataset]: Set5, Set14, B100, Urban100
Example:
python test.py -v "ASID_X2_DIV2K" -t tetser_Matlab -s 0 --test_dataset_name Set5
- Execute ./PSNR_SSIM_Evaluate.m for PSNR/SSIM report. Make sure the location configuration and scale are correct in the Matlab file.
-
Prepare training dataset (DIV2K) from here, then set the location configuration in
./env/env.json -
Prepare yaml file which contains training details in
./train_yamls/ -
Train the models with the following cmd
[Training from scratch] python train.py -v "Model_Name" -p train --train_yaml "[Training_Setting.yaml]"
[Finetuning] python train.py -v "Model_Name" -p finetune -e [Epoch Number of the Pretrained Model] --train_yaml "[Training_Setting.yaml]"
Example:
python train.py -v "ASID_fromscratch_X2_DIV2K" -p train --train_yaml "train_ASID_X2_DIV2K.yaml"
python train.py -v "ASID_fromscratch_X2_DIV2K" -p finetune -e 500 --train_yaml "train_ASIDfinetune_X2_DIV2K.yaml"
Our model was trained from scratch using a learning rate of 5e-4 for 500 epochs, and then fine-tuned with a learning rate of 4e-4 for an additional 1000 epochs.
@inproceedings{park2025efficient,
title={Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution},
author={Park, Karam and Soh, Jae Woong and Cho, Nam Ik},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={},
number={},
pages={},
year={2025}
}