MSFS-Net

Multi-scale frequency separation network for image deblurring

Installation

Python 3.7.13
pytorch  1.9.0
CUDA 10.2
scikit-image
opencv-python
Tensorboard

Pretrained Models

We provide our pre-trained models. You can test our network according to the instruction below.

Baidu link：https://pan.baidu.com/s/1FwHEuyivhCP_BynZC0Ayjw password：0516

Google drive：

https://drive.google.com/drive/folders/1l0A8l1zqJJ6KOqNizQSFQIH3tksjOUMt?usp=sharing

weights	training dataset
model.pkl	GoPro
model_R.pkl	RealBlur-R
model_J.pkl	RealBlur-J

Dataset

prepare datasets

GoPro

• Download deblur dataset from the GoPro dataset.

• Unzip files dataset folder.

• Preprocess dataset by running the command below:

  python data/preprocessing.py

• After preparing data set, the data folder should be like the format below:

  GOPRO
  ├─ train
  │ ├─ blur    % 2103 image pairs
  │ │ ├─ xxxx.png
  │ │ ├─ ......
  │ │
  │ ├─ sharp
  │ │ ├─ xxxx.png
  │ │ ├─ ......
  │
  ├─ test    % 1111 image pairs
  │ ├─ ...... (same as train)
  
  

HIDE

• Download deblur dataset from the HIDE dataset

• Preprocess dataset by running the command below:

  python data/HIDE.py

  note: Please change the path of the dataset location in the code

  format：the same as GoPro datasets

RealBlur

• Download deblur dataset from the RealBlur dataset
• The data folder should be like the format of GoPro datasets.

Test

GoPro and HIDE

To test MSFS-Net，run the command below:

python main.py --model_name "MSFS-Net" --mode "test" --data_dir "dataset/GOPRO" --test_model "model.pkl"

note:You should change line 32 of main.py to model=build_net()

or to test MSFS-Net-Local, run the command below:

python main.py --model_name "MSFS-Net-Local" --mode "test" --data_dir "dataset/GOPRO" --test_model "model.pkl"

note:You should change line 32 of main.py to model=build_arch_net()

Output images will be saved in results/model_name/result_image folder.

RealBlur

The run command is the same as above，but you should change line 15 of main.py to from eval_R import _eval

PSNR and SSIM

We measured PSNR and SSIM using matlab functions.