Lingchen Sun^1,2 | Jie Liang² | Shuaizheng Liu^1,2 | Hongwei Yong¹ | Lei Zhang^1,2

¹The Hong Kong Polytechnic University, ²OPPO Research Institute

⭐ If EA-Adam is helpful to your images or projects, please help star this repo. Thanks! 🤗

## git clone this repository
git clone https://github.com/csslc/EA-Adam.git
cd EA-Adam

# create an environment
conda create -n EA-Adam python=3.10 -y
conda activate EA-Adam
pip install -r requirements.txt

Download from GoogleDrive.

Download from BaiduNetdisk (pwd: 0930).

You can put the testing images in the test_input.

python test.py \
--input_image test_input \
--config configs/mulsrresnet_gan.yml \
--test_model_path pretrained_models/EA-Adam-srresnet.pt \
--root_img output

If you want to test RRDB-based and SwinIR-based models, please modify the test_model_path and config accordingly.

We take the SRResNet backbone as example. Please check and adapt the config files firstly.

1. EA-Adam stage.

A model pretrained with L1 loss needs to be used for stable training, similar to SRGAN, ESRGAN, and other GAN-based SR models.

python train_EA-Adam_srresnet.py \
--config configs/mulsrresnet_gan.yml \
--resume l1-pretrained/ \

2. Weight regression network training.

N expert models can be obtained during the EA-Adam stage and should be placed in fusion_experts for training weight regression network. The final discriminator from the EA-Adam stage can serve as a pretrained model to facilitate faster convergence.

python train_fusion_srresnet.py \
--config configs/fusion_srresnet.yml \
--expert_path fusion_experts/ \
--resume_d disc.pt

3. Model fusion.

The fusion models from weight regression network training stage is placed in fusion_model. The final model is saved in final_model_path.

python train_cal_weight.py \
--config configs/fusion_srresnet.yml \
--test_model_path fusion_model/ \
--expert_path fusion_experts/ \
--final_model_path experiments/EA-Adam-srresnet/ 

This project is released under the Apache 2.0 license.

This project is built based on the SimpleIR and E-GAN projects. Thanks for their awesome works.

If our code helps your research or work, please consider citing our paper. The following are BibTeX references:

@article{sun2024eaadam,
  title={Perception-Distortion Balanced Super-Resolution: A Multi-Objective Optimization Perspective},
  author={Sun, Lingchen and Liang, Jie and Liu, Shuaizheng and Yong, Hongwei and Zhang, Lei},
  journal={IEEE Transactions on Image Processing},
  year={2024}
}

If you have any questions, please feel free to contact: [email protected]