zxw-king/lsnet

source code of "LSNet: Extremely lightweight Siamese Network for Change Detection in Remote Sensing Images"

LSNet

Official implementation of "LSNet: Extremely lightweight Siamese Network for Change Detection in Remote Sensing Images". arxiv|IGARSS2022

Requirements

We use python/pytorch/torchvision versions as follow:

• python 3.7
• pytorch 1.10.0
• torchvision 0.11.1

You can try a lower version, but the python version is no less than 3.6 and the pytorch version is not less than 1.5. If you have any questions, please submit issue.

Dataset

We use CDD dataset from Change Detection in Remote Sensing Images Using Conditional Adversarial Networks

Train

For training, you can modify parameters in "metadata.json", or just keep the default and:

python train.py

Test

All the pre-trained models have been upload in ./weights, you can modify "model name" in "metadata.json" and

python eval.py

Noted that the source code has been reconstructed and the results are a little different from the paper. But still keeping efficient.

Citation

If you feel it useful, please star and cite our work:

@inproceedings{liu2022lsnet,
  title={LSNET: Extremely Light-Weight Siamese Network for Change Detection of Remote Sensing Image},
  author={Liu, Biyuan and Chen, Huaixin and Wang, Zhixi and Xie, Wenqiang and Shuai, LingYu},
  booktitle={IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium},
  pages={2358--2361},
  year={2022},
  organization={IEEE}
}

References

Note that the source code is implemented with reference to SNUNet.