feiyuli-cs/ffnet

Joint Demosaicing and Denoising with Frequency Domain Features

FFNet

Joint Demosaicing and Denoising with Frequency Domain Features

Contents

1. Environment
2. Train
3. Test
4. Dataset
5. Other

Environment

python=3.8 numpy=1.21.2 opencv-python=4.5.5.64
pillow=8.4.0 numba=0.55.1 scikit-image=0.18.3
pytorch=1.10.0 torchvision=0.11.1 cudatoolkit=11.3

Train

train JDD:

python train.py --phase train --task JDnDm --model FFNet --in_type noisy_rgb

train DM:

python train.py --phase train --task DM --model FFNet-DM-B --in_type rgb

If you want to train other model, please change --model "your model name". The model weights will be saved in ./logs/.../checkpoint/xxx.pth folder.

Test

Our pretrain models in Google drive.

For JDD:

To test FFNet, run the command below:

python jdd_test.py --phase test --model FFNet --test_path dataset/JDnDm/test/ --pretrain logs/JDnDm/DIV2K/model_FFNet-in_type_noisy_rgb-C_64-B_32-Patch_128-Epoch_200/checkpoint/model_FFNet-in_type_noisy_rgb-C_64-B_32-Patch_128-Epoch_200_checkpoint_best.pth

For DM:

To test FFNet-DM-B, run the command below:

python dm_test.py --phase test --model FFNet-DM-B --test_path dataset/JDnDm/test/ --pretrain logs/JDnDm/DIV2K/model_FFNet-DM-B-in_type_rgb-C_64-B_32-Patch_128-Epoch_200/checkpoint/model_FFNet-DM-B-in_type_rgb-C_64-B_32-Patch_128-Epoch_200_checkpoint_best.pth 

The test logs will be saved in ./logs/.../xxx_test.log folder and results will be saved in ./logs/.../results/... folder.

Dataset

Download DIV2K train dataset and Kodak, McMaster and Urban100 test datasets (you can also download test datasets in our Google Drive).

You can obtain the train subdataset from BasicSR by scipts/data_preparation/extract_subimages.py (only extract HR images).

The data folder should be like the format below:

dataset
├─ DIV2K
│ ├─ train     % 32592 images
│ │ ├─ DIV2K_train_HR_sub
│ │ |  ├─ xxxx.png
│ │ |  ├─ ...
│ | |
| | |
│ ├─ valid     % 4152 images
│ │ ├─ DIV2K_valid_HR_sub
│ │ |  ├─ xxxx.png
│ │ |  ├─ ...
|
|
├─ JDnDM
│ ├─ test
| │ ├─ Kodak     
| │ │ ├─ xxxx.png
│ | | ├─ ......
│ │ |
| | |
│ | ├─ McMaster
│ │ | ├─ xxxx.png
│ │ | ├─ ......
| | |
| | |
│ | ├─ Urban100   
│ │ | ├─ xxxx.png
│ │ | ├─ ......
...

Other

For JDD: JDNDMSR, CDLNet.

For DM: IRCNN, DPIR, RSTCANet.