Pretrained Weight is released here. This is the checkpoint originally used to generate the submission in NTIRE'20. Specific keys might be different because the change in pytorch lightning, but the weights should work
Moire patterns are commonly seen artifacts when taking photos of screens and other objects with high-frequency textures. It's challenging to remove the moire patterns considering its complex color and shape. In this work, we propose an Attentive Fractal Network to effectively solve this problem. First, we construct each Attentive Fractal Block with progressive feature fusion and channel-wise attention guidance. The network is then fractally stacked with the block on each of its levels. Second, to further boost the performance, we adopt a two-stage augmented refinement strategy. With these designs, our method wins the burst demoireing track and achieves second place in single image demoireing and single image deblurring tracks in NTIRE20 Challenges. Extensive experiments demonstrate the superiority of our method for moire pattern removal compared to existing state-of-the-art methods, and prove the effectiveness of its each component.
We use the original datasets released in the NTIRE20 demoire and deblur challenges, please visit the challenge page or register on the corresponding Codalab competitions to obtain the datasets.
create and activate a virtual environment with Python >= 3.6
Install requirements
pip install -r requirements.txtThere are a few code you must modify before training.
-
Modify variable
TR_INPUTandTR_GTindata.py. TR_INPUT indicates the input path for the whole released training set TR_GT indicates the gt path for the whole released training set -
(optional) Dump a numpy.ndarray for indexing train set and val set with ndim==1 into train.pkl and val.pkl
-
Run command below to start training.
python main.py -M AFN --batch_size <BSZ> --lr <initial value used in cos lr> --max_epochs <n_epoch>
You may want to check full command line arguments available with --help.
-
Set
TS_INPUTindata.pyto the path of result for testing. -
Run command below to start inferring.
python main.py --mode test -M AFN -C <checkpoint path> --timestamp final_resu_test
Results will be located in
res/final_resu_test
You can add new models in models/ folder and use -M to specify the used model.
You can add new loss function in loss/ folder and use -L to specify the used loss function.
If you find this code helpful for your research, please cite our paper Moire Removal via Attentive Fractal Network:
@InProceedings{Xu_2020_CVPR_Workshops,
author = {Xu, Dejia and Chu, Yihao and Sun, Qingyan},
title = {Moire Pattern Removal via Attentive Fractal Network},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2020}
}
By OIerM team.