Intriguing Findings of Frequency Selection for Image Deblurring (AAAI 2023 Oral)

Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen and Yan Wang

Paper: https://arxiv.org/abs/2111.11745 (v2)

code: https://github.com/DeepMed-Lab-ECNU/DeepRFT-AAAI2023

Deep Residual Fourier Transformation for Single Image Deblurring

Xintian Mao, Yiming Liu, Wei Shen, Qingli Li and Yan Wang

Paper: https://arxiv.org/abs/2111.11745v1 (v1)

Network Architecture

Overall Framework of DeepRFT

Installation

The model is built in PyTorch 1.8.0 and tested on Ubuntu 18.04 environment (Python3.8, CUDA11.1).

For installing, follow these intructions

conda create -n pytorch python=3.8
conda activate pytorch
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install matplotlib scikit-image opencv-python yacs joblib natsort h5py tqdm kornia tensorboard ptflops

Install warmup scheduler

cd pytorch-gradual-warmup-lr; python setup.py install; cd ..

Quick Run (If you meet the performance drop, please try branch-AAAI2023 or train the model by yourself.)

To test the pre-trained models of Deblur and Defocus Google Drive or 百度网盘(提取码:phws) on your own images, run

python test.py --weights ckpt_path_here --input_dir path_to_images --result_dir save_images_here --win_size 256 --num_res 8 [4:small, 20:plus]# deblur
python test.py --weights ckpt_path_here --input_dir path_to_images --result_dir save_images_here --win_size 512 --num_res 8 # defocus

Here is an example to train:

python train.py

Results

Experiment for image deblurring.

Deblurring on GoPro Datasets.

Reference Code:

• https://github.com/yangyanli/DO-Conv
• https://github.com/swz30/MPRNet
• https://github.com/chosj95/MIMO-UNet
• https://github.com/codeslake/IFAN

Citation

If you use DeepRFT, please consider citing:

@inproceedings{xintm2023DeepRFTv2, 
    title = {Intriguing Findings of Frequency Selection for Image Deblurring},
    author = {Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen and Yan Wang}, 
    booktitle = {Proceedings of the 37th AAAI Conference on Artificial Intelligence}, 
    year = {2023}
    }

or

@inproceedings{xintm2023DeepRFTv1,
    title={Deep Residual Fourier Transformation for Single Image Deblurring},
    author={Xintian Mao, Yiming Liu, Wei Shen, Qingli Li, Yan Wang},
    booktitle={arXiv:2111.11745},
    year={2021}
}

Contact

If you have any question, please contact mxt_invoker1997@163.com

Our Related Works

• LoFormer: Local Frequency Transformer for Image Deblurring, ACM MM 2024. Paper | Code
• AdaRevD: Adaptive Patch Exiting Reversible Decoder Pushes the Limit of Image Deblurring, CVPR 2024. Paper | Code