Remote Sensing Image Super-Resolution via Saliency-Guided Feedback GANs

by Hanlin Wu, Libao Zhang, and Jie Ma, details are in paper.

Introduction

This repository is build for the proposed SG-FBGAN, which contains full training and testing code.

Usage

Clone the repository:

git clone https://github.com/BNUAI/SG-FBGAN.git

Requirement:

• tensorflow==1.14.0
• tensorlayer==1.11.0
• numpy
• easydict
• opencv-python
• tqdm
• wget

pip install -r requirements.txt

Test with our pre-trained models:

1. Download the pre-trained SG-FBGAN models.

• BI degradation
  • x2: BI_x2.zip
  • x3: BI_x3.zip
  • x4: BI_x4.zip
  • x8: BI_x8.zip
• DN degradation
  • x2: DN_x2.zip
  • x3: DN_x3.zip
  • x4: DN_x4.zip
  • x8: DN_x8.zip

1. Unzip the the downloaded file, and put the pre-trained model on path : experiments/exp_name
2. Do testing:

   python predict.py --opt exp_name
   

   Note: The GeoEye-1 dataset will be downloaded automatically. If the download fails, please download it manually from here, and then put the downloaded file on path : data/sr_geo.npz.

Train with our GeoEye dataset:

python train.py --opt config/va_fbgan_x3_BI.json

Train with your own dataset:

1. change the datapath and savepath in data_loader/make_npz.py, and then make the .npz file:

   python data_loader/make_npz.py
   

2. change the data_path in config/your_own_config_file.json.

3. Do training:

   python train.py --opt config/your_own_config_file.json
   

Cite

H. Wu, L. Zhang and J. Ma, "Remote Sensing Image Super-Resolution via Saliency-Guided Feedback GANs," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2020.3042515.

Contact

• Hanlin Wu ([email protected])