HanzhouLiu / DeblurDiNAT

Official implementation of the paper "DeblurDiNAT: A Lightweight and Effective Transformer for Image Deblurring".

DeblurDiNAT: A Lightweight and Effective Transformer for Image Deblurring

Hanzhou Liu, Binghan Li, Chengkai Liu, Mi Lu

This is the Official Pytorch Implementation of DeblurDiNAT.

Visual Results

Blurry	DeblurDiNAT-L	FFTformer	Uformer-B	Stripformer	Restormer

Quantitative Results

Installation

The implementation is modified from "DeblurGANv2".

git clone https://github.com/HanzhouLiu/DeblurDiNAT.git
cd DeblurDiNAT
conda create -n DeblurDiNAT python=3.8
source activate DeblurDiNAT
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
pip install opencv-python tqdm pyyaml joblib glog scikit-image tensorboardX albumentations
pip install -U albumentations[imgaug]
pip install albumentations==1.1.0

The NATTEN package is required. Please follow the NATTEN installation instructions "NATTEN Homepage". Make sure Python, PyTorch, and CUDA versions are compatible with NATTEN.

Training

Download "GoPro" dataset into './datasets'
For example: './datasets/GoPro'

We train our DeblurDiNAT in two stages:

• We pre-train DeblurDiNAT for 3000 epochs with patch size 256x256
  
• Run the following command

python train_DeblurDiNAT_pretrained.py

• After 3000 epochs, we keep training DeblurDiNAT for 1000 epochs with patch size 512x512
  
• Run the following command

python train_DeblurDiNAT_gopro.py

Testing

For reproducing our results on GoPro and HIDE datasets, download "DeblurDiNATL.pth"

For testing on GoPro dataset

• Download "GoPro" full dataset or test set into './datasets' (For example: './datasets/GoPro/test')
  
• Run the following command

python predict_GoPro_test_results.py --job_name DeblurDiNATL --weight_name DeblurDiNATL.pth --blur_path ./datasets/GOPRO/test/blur

For testing on HIDE dataset

• Download "HIDE" into './datasets'
  
• Run the following command

python predict_HIDE_results.py --weights_path ./DeblurDiNATL.pth 

For testing on RealBlur test sets

• Download "RealBlur_J" and "RealBlur_R" into './datasets'
  
• Run the following command

python predict_RealBlur_J_test_results.py --weights_path ./DeblurDiNATL.pth 

python predict_RealBlur_R_test_results.py --weights_path ./DeblurDiNATL.pth 

Citation

@misc{liu2024deblurdinat,
      title={DeblurDiNAT: A Lightweight and Effective Transformer for Image Deblurring}, 
      author={Hanzhou Liu and Binghan Li and Chengkai Liu and Mi Lu},
      year={2024},
      eprint={2403.13163},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}