Authors: Zihao Xiong, Fei Zhou, Fengyi Wu, Shuai Yuan, Maixia Fu, Zhenming Peng, Jian Yang, Yimian Dai
Official implementation of "DRPCA-Net: Make Robust PCA Great Again for Infrared Small Target Detection".

• Abstract
• Installation
• Dataset Preparation
• Training
• Testing
• Model Zoo and Benchmark
• Citation
• Acknowledgements

Infrared small target detection plays a critical role in remote sensing, industrial monitoring, and civilian applications.

Despite the progress of deep learning methods, many end-to-end models sacrifice interpretability and generalization. These models often ignore the intrinsic sparsity prior of infrared small targets — a fundamental cue.

DRPCA-Net rethinks this by:

• Revisiting traditional Robust PCA with deep learning.
• Introducing Dynamic Unfolding through a lightweight hypernetwork.
• Proposing a Dynamic Residual Group (DRG) to handle complex backgrounds.

Outcome: Superior robustness, generalization, and efficiency, validated across multiple public datasets.

git clone https://github.com/GrokCV/DRPCA-Net.git
cd DRPCA-Net

conda create -n drpcanet python=3.9
conda activate drpcanet
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt

We used SIRSTV1, NUDT-SIRST, IRSTD-1K and SIRST-Aug for both training and test.

• NUDT-SIRST can be found and downloaded here: NUDT-SIRST
• IRSTD-1K can be found and downloaded here: IRSTD-1K
• SIRST-Aug can be found and downloaded here: SIRST-Aug
• SIRSTV1 can be found and downloaded here: SIRSTV1

Note:

You can also refer to Google Drive. - Wu Fengyi's settings for ready-to-use datasets.

For the SIRSTV1 dataset, we still maintain the structure provided by the official source.

Please first download these datasets and place NUDT-SIRST, IRSTD-1K, and sirst_aug in the folder ./datasets/; for the SIRSTV1 dataset, please maintain its official directory structure.

├──./datasets/
│    ├── NUDT-SIRST/ & IRSTD-1k/ & sirst_aug/
│    │    ├── trainval
│    │    │    ├── images
│    │    │    │    ├── 000002.png
│    │    │    │    ├── 000004.png
│    │    │    │    ├── ...
│    │    │    ├── masks
│    │    │    │    ├── 000002.png
│    │    │    │    ├── 000004.png
│    │    │    │    ├── ...
│    │    ├── test
│    │    │    ├── images
│    │    │    │    ├── 000001.png
│    │    │    │    ├── 000003.png
│    │    │    │    ├── ...
│    │    │    ├── masks
│    │    │    │    ├── 000001.png
│    │    │    │    ├── 000003.png
│    │    │    │    ├── ...

Please make sure that the path of your data set is consistent with the data_root.

Runrun_configs.py to perform network training:

$ python run_configs.py

In the run_configs.py file, by setting parameters such as data_irstd1k, data_sirstaug, data_nudt, and SirstDataset, you can respectively run the corresponding datasets for training.

Run inference.py to generate file of the format .mat and .png:

$ python inference.py

In the inference.py script, you can perform inference on different datasets by specifying the dataset_name parameter with one of the following options: NUDT-SIRST, IRSTD-1k, sirst_aug, or SIRSTv1. Each value corresponds to a specific dataset configuration.

Result files (.png and .mat) will be saved to:

├──./results/
│    ├── [dataset_name]
│    │   ├── img
│    │   │    ├── 000000.png
│    │   │    ├── 000001.png
│    │   │    ├── ...
│    │   ├── mat
│    │   │    ├── 000000.mat
│    │   │    ├── 000001.mat
│    │   │    ├── ...

Run cal_from_mask.py for direct evaluation:

$ python cal_from_mask.py

In the cal_from_mask.py file, you can calculate metrics such as mIoU, F1, Pd, and Fa for the inference results of corresponding datasets by setting the data_list parameter to one of the following values: NUDT-SIRST, IRSTD-1k, sirst_aug, or SIRSTv1.

The checkpoints folder contains the weights for four datasets.

@ARTICLE{TGRS2025DRPCANet,
  author={Xiong, Zihao and Zhou, Fei and Wu, Fengyi and Yuan, Shuai and Fu, Maixia and Peng, Zhenming and Yang, Jian and Dai, Yimian},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={DRPCA-Net: Make Robust PCA Great Again for Infrared Small Target Detection}, 
  year={2025},
  volume={63},
  number={},
  pages={1-16},
  doi={10.1109/TGRS.2025.3588392}}

🔥 Let's Make Robust PCA Great Again!