This is an unofficial reproduction of paper Moir´e Photo Restoration Using Multiresolution Convolutional Neural Networks.(PyTorch)
First of all, you need to prepare the whole dataset of this paper, which is around 100G.
dataset download link : https://drive.google.com/drive/folders/109cAIZ0ffKLt34P7hOMKUO14j3gww2UC
another download link : https://pan.baidu.com/s/16MWtsIqDueaBwR45AuPMMA (code:h3k9)
- torch >= 1.6.0
- torchvision >= 0.7.0
- pillow >= 7.2.0
- GPU >= 3G
Before starting to train the model, you need to run a script to clean the training set as shown below.
All hyper-parameters follow the instructions of the paper, so you don't need to change them.W
However, you should change the path of datasets to match your local environment.
python utils.py
python train.py --dataset /data_new/zxbsmk/moire/trainData --save ./model
Get PSNR of the testing set.
python test.py
psnr distribution | <12 | 12~14 | 14~17 | 17~20 | 20~22 | 22~24 | >24 |
---|---|---|---|---|---|---|---|
training set | 72 | 2318 | 29816 | 37089 | 21195 | 15102 | 12856 |
testing set | 8 | 227 | 2951 | 3809 | 2069 | 1463 | 1324 |
total | 80 | 2545 | 32767 | 40898 | 23264 | 16565 | 14180 |
We can see that low quality image pairs whose PSNR is lower than 12 still exist in the dataset, which is against the author's declaration.