Extend-GAN

This work can be used to extend the boundaries of a high-resolution image to the extent of a given low-resolution reference image:

Environment

Tested on Ubuntu 20.04. Python version 3.10. Pytorch version 1.13.1.

Create your environment using this command:

mamba env create -f environment_1.13.1.yaml

If you use conda, replace mamba with conda.

Dataset

Prepare your data and use util/crop_rs.py to crop HR and corresponding LR images.

You will get folders following this structure:

dataset
├─train
│  ├─source
│  │      source1.tif
│  │      source2.tif
│  │      ...
│  │
│  ├─ref
│  │      ref1.tif
│  │      ref2.tif
│  ├...
│  
└─test
   ├─source
   │      source1.tif
   │      source2.tif
   │      ...
   │
   ├─ref
   │      ref1.tif
   │      ref2.tif
   └─...

Generate flists in current directory.

ls -R ${YOUR_ABSOLUTE_PATH} > ${FLIST_NAME}
# for example
ls -R /data/cyc/dataset/train/source/*.tif > train.flist
ls -R /data/cyc/dataset/test/source/*.tif > test.flist

The default size is 512, use util/crop_256.ipynb to randomly crop images, if necessary.

Train

python train.py --batch_size ${BATCH_SIZE} --train_dataset_name ${YOUR_TRAIN_FLIST} --n_epochs ${TOTAL_EPOCHS}
# for example
python train.py --batch_size 8 --train_dataset_name /data/cyc/dataset/train.flist --n_epochs 2400 > log_42.txt 

Test

python test.py --image_path ${YOUR_TEST_FLIST} --model ${YOUR_GENERATOR_PATH}
# for example
python test.py --image_path /data/cyc/dataset/test.flist --model saved_models/generator_2400.pth

Acknowledgments

We are benefiting a lot from the following projects:

• Image-Inpainting-Implementations

• Boundless-in-Pytorch

• Palette-Image-to-Image-Diffusion-Models

If you find this work useful, please cite:

@ARTICLE{10413911,
  author={Cui, Yongchuan and Liu, Peng and Song, Bingze and Zhao, Lingjun and Ma, Yan and Chen, Lajiao},
  journal={IEEE Geoscience and Remote Sensing Letters}, 
  title={Reconstruction of Large-Scale Missing Data in Remote Sensing Images Using Extend-GAN}, 
  year={2024},
  volume={21},
  number={},
  pages={1-5},
  keywords={Training;Earth;Artificial satellites;Generative adversarial networks;Spatial resolution;Remote sensing;Image reconstruction;Generative adversarial network (GAN);image reconstruction;remote sensing images;triplet loss},
  doi={10.1109/LGRS.2023.3317898}}