/MRDA

This project is the official implementation of 'Meta-Learning based Degradation Representation for Blind Super-Resolution', TIP2023

Primary LanguagePython

MRDA

This project is the official implementation of 'Meta-Learning based Degradation Representation for Blind Super-Resolution', TIP2023

Meta-Learning based Degradation Representation for Blind Super-Resolution [Paper] [Project]

This is code for MRDA (for classic degradation model, ie y=kx+n)

Dependencies and Installation

Dataset Preparation

We use DF2K, which combines DIV2K (800 images) and Flickr2K (2650 images).

Training

  1. train Meta-Learning Network (MLN) bicubic pretraining
sh main_stage1.sh

Isotropic Gaussian Kernels

  1. we train MLN using meta-learning scheme. It is notable that modify the ''pre_train'' of main_stage2.sh to the path of trained main_stage1 checkpoint. Then, we run
sh main_stage2.sh
  1. we train MLN with teacher MRDA_T together. It is notable that modify the ''pre_train_meta'' of main_stage3.sh to the path of trained main_stage2 checkpoint. Then, we run
sh main_stage3.sh
  1. we train student MRDA_S. It is notable that modify the ''pre_train_meta'' of main_stage3.sh to the path of trained main_stage2 checkpoint. ''pre_train_TA'' and ''pre_train_ST'' are both set to the path of trained main_stage3 checkpoint.. Then, we run
sh main_stage4.sh

Anisotropic Gaussian Kernels plus noise

It's training process is the same as isotropic Gaussian Kernels, except we use the anisotropic Gaussian Kernels settings in main_stage2.sh, main_stage3.sh, and main_stage4.sh .

  1. we train MLN using meta-learning scheme. It is notable that modify the ''pre_train'' of main_stage2.sh to the path of trained main_stage1 checkpoint. Then, we run
sh main_stage2.sh
  1. we train MLN with teacher MRDA_T together. It is notable that modify the ''pre_train_meta'' of main_stage3.sh to the path of trained main_stage2 checkpoint. Then, we run
sh main_stage3.sh
  1. we train student MRDA_S. It is notable that modify the ''pre_train_meta'' of main_stage3.sh to the path of trained main_stage2 checkpoint. ''pre_train_TA'' and ''pre_train_ST'' are both set to the path of trained main_stage3 checkpoint.. Then, we run
sh main_stage4.sh

🏰 Model Zoo

Please download checkpoints from Google Drive.

Testing

Isotropic Gaussian Kernels

sh test_iso_stage4.sh

Anisotropic Gaussian Kernels plus noise

sh test_anisoAnoise_stage4.sh

Results

BibTeX

@article{xia2022meta,
  title={Meta-learning based degradation representation for blind super-resolution},
  author={Xia, Bin and Tian, Yapeng and Zhang, Yulun and Hang, Yucheng and Yang, Wenming and Liao, Qingmin},
  journal={IEEE Transactions on Image Processing},
  year={2023}
}

📧 Contact

If you have any question, please email zjbinxia@gmail.com.