Few-shot class incremental learning for efficient SAR-ATR

Our paper #Few-Shot Class-Incremental Learning For Efficient SAR Automatic Target Recognition has been accepted for publication in IEEE ICIP 2025.

The implementation of IEEE ICIP 2025 paper Few-Shot Class-Incremental Learning For Efficient SAR Automatic Target Recognition (You can also read our paper here.)

If you use any code of this repo, please consider citing our work:

@article{karantaidis2025few,
  title={Few-Shot Class-Incremental Learning For Efficient SAR Automatic Target Recognition},
  author={Karantaidis, G. and Pantsios, A. and Kompatsiaris, I. and Papadopoulos, S.},
  journal={arXiv preprint arXiv:2505.19565},
  year={2025}
}

Install

conda create -n DILHyFS python=3.9
conda activate DILHyFS 
pip install -r requirements.txt

Dependencies

This code is implemented in PyTorch, and we perform the experiments under the following environment settings:

torch==2.0.1
torchvision==0.15.2
scikit_learn
scipy
tqdm
numpy==1.26.4

Dataset Preparation

• Create a folder "datasets/" under the root directory
• MSTAR: download the folders MSTAR from link and place them into the 'datasets/' folder
• MSTAR_OPENSAR: download the folder MSTAR_OPENSAR from link and place them into the 'datasets/' folder.
• SAR-AIRcraft-1.0: download the folder AIRCRAFT link and place them into the 'datasets/' folder.

Training:

• To train DILHyFS in the FSCIL scenario, place the pre-trained weights of the GFNet model link into the 'weights/' folder.
• For MSTAR dataset run:

python main.py --config=./exps/DILHyFS.json

• For cross-domain experiment run:

python main.py --config=./exps/cross_exp.json

• For the comparative analysis of the backbone networks run:

python main.py --config=./exps/ablation_exps/resnet.json
python main.py --config=./exps/ablation_exps/gfnet.json

• For DILHyFS evaluation on limited data scenario run:

python main.py --config=./exps/limited.json

Acknowledgments

We thank the following repos providing helpful components/functions in our work.

• PILOT
• RanPAC
• GFNet