chu246 / msprl

Rethinking PRL: A Multiscale Progressively Residual Learning Network for Inverse Halftoning

MS-PRL

Rethinking PRL: A Multiscale Progressively Residual Learning Network for Inverse Halftoning.

Abstract

Network architecture

Contents

1. Environment
2. Demo
3. Train
4. Test and Valid
5. Dataset
6. Model
7. Citation
8. 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

Demo

demo images in demo/halftone/ folder, and the output images in demo/output/ folder.

python demo.py

Train

To train MS-PRL , run the command below:

python main.py --mode train --model_name=MSPRL

if you want to train other model, pleace change --model_name="your model name". The model weights will be saved in ./checkpoint/model_name/model_name_iterations.pth folder.

Test and Valid

1. run test mode, images will be saved in ./resutls/model_name/test_name/ and the log will be saved in ./logs/model_name/test/test_name/log.txt.
2. run valid mode, just the log will be saved in ./logs/model_name/test/test_name/log.txt

To test MS-PRL , run the command below:

python main.py --mode test --model_name=MS-PRL

To valid MS-PRL , run the command below:

python main.py --mode valid --model_name=MS-PRL

Please pay attention to the dataset path, refer to the details of the dataset.

Dataset

1. Download VOC2012, Kodak25, Place365 dataset and five standard benchmark datasets. You can also download our dataset in here.

2. To generate halftone image using Floyd Steinberg error diffusion, run the command below:

cd utils
python halftone.py

The data folder should be like the format below:

dataset
├─ train
│ ├─ data     % 13841 halftone images
│ │ ├─ xxxx.png
│ │ ├─ ......
│ │
│ ├─ target   % 13841 gray images
│ │ ├─ xxxx.png
│ │ ├─ ......
│
├─ valid
│ ├─ data     % 3000 halftone images
│ │ ├─ xxxx.png
│ │ ├─ ......
│ │
│ ├─ target   % 3000 gray images
│ │ ├─ xxxx.png
│ │ ├─ ......
|
├─ test
│ ├─ Class
| │ ├─ data     % halftone images
| │ │ ├─ xxxx.png
│ | │ ├─ ......
│ │
│ | ├─ target   % gray image
│ │ | ├─ xxxx.png
│ │ | ├─ ......
|
│ ├─ Kodak
| | ├─ ......

Model

We provide our all pre-trained models.

• MS-PRL, PRL-dt and other model in here. The data folder should be like the format below:

checkpoint
├─ MSPRL
│ ├─ MSPRL_iteration.pth
│ │
├─ DnCNN
│ ├─ DnCNN_iteration.pth
│ │

Citation

Other

Reference Code:

1. https://github.com/chosj95/MIMO-UNet
2. https://github.com/swz30/MIRNetv2
3. https://github.com/csbhr/CNLRN
4. DnCNN: https://github.com/cszn/KAIR