This repository is the Official Pytorch Implementation of Progressive Semantic Face Deblurring.
T. B. Lee, S. H. Jung and Y. S. Heo, "Progressive Semantic Face Deblurring," in IEEE Access, vol. 8, pp. 223548-223561, 2020, doi: 10.1109/ACCESS.2020.3033890.
- Python >= 3.5
- Pytorch >= 1.2.0
- Download training data and training kerles from here: MSPL_TrainingData Google Drive
- Extract 'train_img.zip' and 'train_label_cls4.zip' into the same folder path you want to specify.
- /path/to/train_data (name is not important)
-- train_img
-- files..
-- train_label_cls4
-- files..
- Extract 'validationset.zip' to the folder path you want to specify (/path/to/valid_data).
- Extract 'trainkernels.zip' to the folder path you want to specify (/path/to/kernel_data).
- Extract 'VGGFace16.zip' to the folder path you want to specify (/path/to/vggface).
- VGGFace16 model weight is came from: Github
- Specify the 'src/confgis/train_configs.py'.
line 10: cfg.GPU_IDS = [0]
- Set the GPU ids you want to training. Default GPU id is 0.
- Set to 'None' to train with CPU.
cfg.GPU_IDS = None.
- To train with multiple GPUs, specify the the GPU device ids.
cfg.GPU_IDS = [1,2] => The model will be trained with two GPUs(device ids(1,2)).
line 14: cfg.SAVE_DIR = '/path/to/save'
- path to save models, training process, etc.
line 37, 41: 'data_dir'
- '/path/to/train_data', '/path/to/valid_data' respectively.
line 48: cfg.TRAIN_KERNEL_PATH = '/path/to/kernel_data'
line 59: 'VGGFace16' = 'Path VGGFace16.pth"- To train the model(s) in the paper, run this command in 'src':
python train.py
- Specify the paths in 'src/inference.py'
MODEL_DIR = '/path/to/trained models/netG_last.pth'
INPUT_IMG_DIR = '/path/to/test image'
OUTPUT_IMG_DIR = '/path/to/save output image'
USE_GPU = 'True of False. When set True, the model will be tested on 'GPU(cuda)' envirionment.'
- To inference the model, run this command in 'src':
python inference.py
- Pretrained model: MSPL_GAN Google Drive
- You can download here: MSPL_Testsets Google Drive
- Test Restuls of ours on MSPL_Testsets Google Drive
- Our model achieves the following performance on MSPL_Testsets :
- PSNR and SSIM values are measured using MATLAB.
| MSPL_Center | PSNR | SSIM | MSPL_Random | PSNR | SSIM |
|---|---|---|---|---|---|
| CelebA | 28.95 | 0.936 | CelebA | 28.07 | 0.921 |
| CelebA-HQ | 29.80 | 0.945 | CelebA-HQ | 28.82 | 0.929 |
| FFHQ | 29.22 | 0.941 | FFHQ | 27.36 | 0.908 |
- Our training and testset are synthesized using CelebA, CelebAMask-HQ, and FFHQ.
- The dataset is available for non-commercial research purposes only.
- You agree not to reproduce, duplicate, copy, sell, trade, resell or exploit for any commercial purposes, any portion of the images and any portion of derived data.
- You agree not to further copy, publish or distribute any portion of the dataset.
- CelebA
- CelebAMask-HQ
- FFHQ
- RealBlurrset
- Codes for MTCNN, MobienetArcface and weights came from: HERE
- Training & Testing kernels are synthesized using the method of,
- Chakrabarti, Ayan. "A neural approach to blind motion deblurring." European conference on computer vision. Springer, Cham, 2016.
- Xia, Zhihao, and Ayan Chakrabarti. "Training Image Estimators without Image Ground Truth." Advances in Neural Information Processing Systems. 2019.
- Shen, Ziyi, et al. "Deep semantic face deblurring." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
- Lu, Boyu, Jun-Cheng Chen, and Rama Chellappa. "Unsupervised domain-specific deblurring via disentangled representations." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.
- Yasarla, Rajeev, Federico Perazzi, and Vishal M. Patel. "Deblurring face images using uncertainty guided multi-stream semantic networks." IEEE Transactions on Image Processing 29 (2020): 6251-6263.
@ARTICLE{lee2020progressive,
author={T. B. {Lee} and S. H. {Jung} and Y. S. {Heo}},
journal={IEEE Access},
title={Progressive Semantic Face Deblurring},
year={2020},
volume={8},
number={},
pages={223548-223561},
doi={10.1109/ACCESS.2020.3033890}}