zhengjun-ai/moirephotorestoration-mcnn

This is an unofficial reproduction of paper Moir´e Photo Restoration Using Multiresolution Convolutional Neural Networks.(PyTorch)

MoirePhotoRestoration-MCNN

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://huggingface.co/datasets/zxbsmk/TIP-2018

Requirements

• torch >= 1.6.0
• torchvision >= 0.7.0
• pillow >= 7.2.0
• GPU >= 3G

Training

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

Testing

Get PSNR of the testing set.

python test.py

Dataset

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.