Realistic Blur Synthesis for Learning Image Deblurring

Project | Paper | Supple

Official Implementation of ECCV Paper

Realistic Blur Synthesis for Learning Image Deblurring
Jaesung Rim, Geonung Kim, Jungeon Kim, Junyong Lee, Seungyong Lee, Sunghyun Cho.
POSTECH
IEEE European Conference on Computer Vision (ECCV) 2022

Pytorch implementation

We provide simple Dataset modules for adopting our pipeline. This is slightly different from the tensorflow implementation.

Please refer to Uformer-RSBlur.

# ./Uformer-RSBlur/dataset/dataset_RealisticDeblur.py

class RealisticGoProABMEDataset(Dataset):
    def __init__(self, image_dir, patch_size=256, image_aug=True, realistic_pipeline=True):
        ...

    def __len__(self):
        return len(self.image_list)

    def __getitem__(self, idx):
        ...

class RealisticGoProUDataset(Dataset):
    def __init__(self, image_dir, patch_size=256, image_aug=True, realistic_pipeline=True):
        ...

    def __len__(self):
        return len(self.image_list)

    def __getitem__(self, idx):
        ...

Results with the proposed method.

Results of analysis (click)

CRF	Interp.	Sat.	Noise	ISP	PSNR / SSIM
Linear					30.12 / 0.7727
sRGB					30.90 / 0.7805
sRGB	✓				30.20 / 0.7468
sRGB	✓		G		31.77 / 0.8275
sRGB	✓	Ours	G		31.83 / 0.8265
sRGB	✓	Ours	G+P	✓	32.06 / 0.8322

Installation

git clone https://github.com/rimchang/RSBlur.git

Tested environment

We recommend a virtual environment using conda or docker.

SRN-Deblur

• Tensorflow 1.15
• cuda11.1

Download

Dataset [Google Drive] [Postech]

Descriptions (click)

• RSBlur
  • 13,358 pairs of real/synthetic blurred image and a corresponding GT image.
• RSBlur_additional
  • 8,821 additional images for learning based synthesis, additional synthetic images or etc.
  • Do not use it as additional real training images.
• RSBlur_sharps
  • All of sharp image sequneces.
• GoPro_INTER_ABME
  • Synthetic blur dataset using GoPro and ABME method.
• GoPro_U
  • Synthetic blur dataset using GoPro and synthetic blur kernels.

RSBlur

# RSBlur.zip

RSBlur
├── 0001
│   ├── 000001
│   │   ├── real_blur/real_blur.png # real blurred image
│   │   ├── avg65_img/avg_blur.png # synthetic blurred image using frame interpolation
│   │   ├── avg65_mask_100/avg_blur.png # saturation mask
│   │   ├── gt/gt_sharp.png # ground truth sharp image
...

GoPro_INTER_ABME

# GoPro_INTER_ABME.zip

GoPro_INTER_ABME
├── GOPR0372_07_00 
│   ├── 000001
│   │   ├── avg_inter_img/avg_blur.png # synthetic blurred image using frame interpolation
│   │   ├── avg_inter_mask_100/avg_blur.png # saturation mask
│   │   ├── gt/gt_sharp.png # ground truth sharp image
...

GoPro_U

# GoPro_U.zip

GoPro_U
├── centroid_blurred_img
│   ├── 0_GOPR0372_07_00_000047_003754_kernel_25_blurred.png # synthetic blurred image using a synthetic blur kernel
│   ...
├── centroid_blurred_mask_100
│   ├── 0_GOPR0372_07_00_000047_003754_kernel_25_blurred.png # saturation mask
│   ...
├── target_img
│   ├── 0_GOPR0372_07_00_000047_003754_kernel_25_gt.png # ground truth sharp image
│   ...

Dataset splits [link]

Pre-trained models [Google Drive] [Postech]

Descriptions (click)

• SRN-Deblur_RSBlur_real : Trained on real set of RSBlur.
• SRN-Deblur_RSBlur_syn_with_ours : Trained on synthetic set of RSBlur with our synthesis pipeline.
• SRN-Deblur_GoPro_ABME_with_ours : Trained on GoPro_INTER_ABME with our synthesis pipeline.
• SRN-Deblur_GoPro_U_with_ours : Trained on GoPro_U with our synthesis pipeline.

Training

# ./SRN-Deblur-RSBlur
# All datasets should be located in SRN-Deblur-RSBlur/dataset

# RSBlur
python run_model.py --phase=train --checkpoint_path=0719_SRN-Deblur_RSBlur_real --sat_synthesis=None --noise_synthesis=None --datalist=../datalist/RSBlur/RSBlur_real_train.txt --gpu=0
python run_model.py --phase=train --checkpoint_path=0719_SRN-Deblur_RSBlur_syn_with_ours --sat_synthesis=sat_synthesis --noise_synthesis=poisson_RSBlur --cam_params_RSBlur=1 --datalist=../datalist/RSBlur/RSBlur_syn_train.txt --gpu=0

# GoPro_INTER_ABME
python run_model.py --phase=train --checkpoint_path=0719_SRN-Deblur_GoPro_ABME_with_ours --target_dataset=RealBlur --sat_synthesis=sat_synthesis --noise_synthesis=poisson_gamma --cam_params_RealBlur=1 --adopt_crf_realblur=1 --datalist=../datalist/GoPro/GoPro_INTER_ABME_train.txt --gpu=0

# GoPro_U
python run_model.py --phase=train --checkpoint_path=0719_SRN-Deblur_U_with_ours --target_dataset=RealBlur --sat_synthesis=sat_synthesis --noise_synthesis=poisson_gamma --cam_params_RealBlur=1 --adopt_crf_realblur=1 --datalist=../datalist/GoPro/GoPro_U_train.txt --gpu=0

Testing

# ./SRN-Deblur-RSBlur
# All datasets should be located in SRN-Deblur-RSBlur/dataset

# RSBlur
python run_model.py --phase=test --checkpoint_path=SRN-Deblur_RSBlur_real --datalist=../datalist/RSBlur/RSBlur_real_test.txt --gpu=0
python run_model.py --phase=test --checkpoint_path=SRN-Deblur_RSBlur_syn_with_ours --datalist=../datalist/RSBlur/RSBlur_real_test.txt --gpu=0

# RealBlur
python run_model.py --phase=test --checkpoint_path=SRN-Deblur_GoPro_ABME_with_ours --datalist=../datalist/RealBlur_J_test_list.txt --gpu=0
python run_model.py --phase=test --checkpoint_path=SRN-Deblur_GoPro_U_with_ours --datalist=../datalist/RealBlur_J_test_list.txt --gpu=0

Evaluation

# ./evaluation

python evaluate_RSBlur.py --input_dir=../SRN-Deblur-RSBlur/testing_res/SRN-Deblur_RSBlur_real --gt_root=../SRN-Deblur-RSBlur/dataset/RSBlur;
python evaluate_RealBlur.py --input_dir=../SRN-Deblur-RSBlur/testing_res/SRN-Deblur_U_with_ours --gt_root=../SRN-Deblur-RSBlur/dataset/RealBlur-J_ECC_IMCORR_centroid_itensity_ref;

Real-world Deblurring Benchmark

We provide an additional deblurring benchmark to provide the basis for future deblurring research. All below models are trained on real blurred images of the RSBlur training set.

Results of benchmark (click)

Methods	PSNR / SSIM	Link
Uformer-B	33.98 / 0.8660	Result / Weight / Test Code
Restormer	33.69 / 0.8628	Result / Weight
MPRNet	33.61 / 0.8614	Result / Weight
MiMO-UNet+	33.37 / 0.8560	Result / Weight
MiMO-UNet	32.73 / 0.8457	Result / Weight
SRN-Deblur	32.53 / 0.8398	Result / Weight

License

The RSBlur dataset is released under CC BY 4.0 license.

Acknowledment

The code is based on SRN-Deblur, CBDNet and UID.

Citation

If you use our dataset for your research, please cite our paper.

@inproceedings{rim_2022_ECCV,
 title={Realistic Blur Synthesis for Learning Image Deblurring},
 author={Rim, Jaesung and Kim, Geonung and Kim, Jungeon and Lee, Junyong and Lee, Seungyong and Cho, Sunghyun},
 booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
 year={2022}
}