Blind2Unblind: Self-Supervised Image Denoising with Visible Blind Spots

Citing Blind2Unblind

@InProceedings{Wang_2022_CVPR,
    author    = {Wang, Zejin and Liu, Jiazheng and Li, Guoqing and Han, Hua},
    title     = {Blind2Unblind: Self-Supervised Image Denoising With Visible Blind Spots},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {2027-2036}
}

Installation

The model is built in Python3.8.5, PyTorch 1.7.1 in Ubuntu 18.04 environment.

Data Preparation

1. Prepare Training Dataset

For processing ImageNet Validation, please run the command
```
python ./dataset_tool.py
```
For processing SIDD Medium Dataset in raw-RGB, please run the command
```
python ./dataset_tool_raw.py
```

2. Prepare Validation Dataset

Please put your dataset under the path: ./Blind2Unblind/data/validation.

Pretrained Models

Download pre-trained models: Google Drive

The pre-trained models are placed in the folder: ./Blind2Unblind/pretrained_models

# # For synthetic denoising
# gauss25
./pretrained_models/g25_112f20_beta19.7.pth
# gauss5_50
./pretrained_models/g5-50_112rf20_beta19.4.pth
# poisson30
./pretrained_models/p30_112f20_beta19.1.pth
# poisson5_50
./pretrained_models/p5-50_112rf20_beta20.pth

# # For raw-RGB denoising
./pretrained_models/rawRGB_112rf20_beta19.4.pth

# # For fluorescence microscopy denoising
# Confocal_FISH
./pretrained_models/Confocal_FISH_112rf20_beta20.pth
# Confocal_MICE
./pretrained_models/Confocal_MICE_112rf20_beta19.7.pth
# TwoPhoton_MICE
./pretrained_models/TwoPhoton_MICE_112rf20_beta20.pth

Train

Train on synthetic dataset

python train_b2u.py --noisetype gauss25 --data_dir ./data/train/Imagenet_val --val_dirs ./data/validation --save_model_path ../experiments/results --log_name b2u_unet_gauss25_112rf20 --Lambda1 1.0 --Lambda2 2.0 --increase_ratio 20.0

Train on SIDD raw-RGB Medium dataset

python train_sidd_b2u.py --data_dir ./data/train/SIDD_Medium_Raw_noisy_sub512 --val_dirs ./data/validation --save_model_path ../experiments/results --log_name b2u_unet_raw_112rf20 --Lambda1 1.0 --Lambda2 2.0 --increase_ratio 20.0

Train on FMDD dataset

python train_fmdd_b2u.py --data_dir ./dataset/fmdd_sub/train --val_dirs ./dataset/fmdd_sub/validation --subfold Confocal_FISH --save_model_path ../experiments/fmdd --log_name Confocal_FISH_b2u_unet_fmdd_112rf20 --Lambda1 1.0 --Lambda2 2.0 --increase_ratio 20.0

Test

Test on Kodak, BSD300 and Set14

For noisetype: gauss25

python test_b2u.py --noisetype gauss25 --checkpoint ./pretrained_models/g25_112f20_beta19.7.pth --test_dirs ./data/validation --save_test_path ./test --log_name b2u_unet_g25_112rf20 --beta 19.7

For noisetype: gauss5_50

python test_b2u.py --noisetype gauss5_50 --checkpoint ./pretrained_models/g5-50_112rf20_beta19.4.pth --test_dirs ./data/validation --save_test_path ./test --log_name b2u_unet_g5_50_112rf20 --beta 19.4

For noisetype: poisson30

python test_b2u.py --noisetype poisson30 --checkpoint ./pretrained_models/p30_112f20_beta19.1.pth --test_dirs ./data/validation --save_test_path ./test --log_name b2u_unet_p30_112rf20 --beta 19.1

For noisetype: poisson5_50

python test_b2u.py --noisetype poisson5_50 --checkpoint ./pretrained_models/p5-50_112rf20_beta20.pth --test_dirs ./data/validation --save_test_path ./test --log_name b2u_unet_p5_50_112rf20 --beta 20.0

Test on SIDD Validation in raw-RGB space

python test_sidd_b2u.py --checkpoint ./pretrained_models/rawRGB_112rf20_beta19.4.pth --test_dirs ./data/validation --save_test_path ./test --log_name validation_b2u_unet_raw_112rf20 --beta 19.4

Test on SIDD Benchmark in raw-RGB space

python benchmark_sidd_b2u.py --checkpoint ./pretrained_models/rawRGB_112rf20_beta19.4.pth --test_dirs ./data/validation --save_test_path ./test --log_name benchmark_b2u_unet_raw_112rf20 --beta 19.4

Test on FMDD Validation

For Confocal_FISH

python test_fmdd_b2u.py --checkpoint ./pretrained_models/Confocal_FISH_112rf20_beta20.pth --test_dirs ./dataset/fmdd_sub/validation --subfold Confocal_FISH --save_test_path ./test --log_name Confocal_FISH_b2u_unet_fmdd_112rf20 --beta 20.0

For Confocal_MICE

python test_fmdd_b2u.py --checkpoint ./pretrained_models/Confocal_MICE_112rf20_beta19.7.pth --test_dirs ./dataset/fmdd_sub/validation --subfold Confocal_MICE --save_test_path ./test --log_name Confocal_MICE_b2u_unet_fmdd_112rf20 --beta 19.7

For TwoPhoton_MICE

python test_fmdd_b2u.py --checkpoint ./pretrained_models/TwoPhoton_MICE_112rf20_beta20.pth --test_dirs ./dataset/fmdd_sub/validation --subfold TwoPhoton_MICE --save_test_path ./test --log_name TwoPhoton_MICE_b2u_unet_fmdd_112rf20 --beta 20.0

zejinwang / Blind2Unblind

readme