README.md

December 25, 2025 · View on GitHub

The official complete code for paper "Multi-Scale Direction-Aware Network for Infrared Small Target Detection" [Paper] [知乎] [Accepted by TGRS2025]

  1. This code integrates a series of infrared small target detection networks we built: MLCL-Net [paper], ALCL-Net [paper], GGL-Net [paper] and MSDA-Net [Paper/arXiv]. For details, please see the "model" folder.

  2. It is worth mentioning that our solution based on MSDA-Net won the first prize in the PRCV2024 Wide-area Infrared Small Target Detection Challenge

  3. Everyone is welcome to follow our new work [PAL Framework] on the task of "Infrared small target detection with single point supervision". On the task of "Infrared small target detection with single point supervision", equipped with our proposed "PAL Framework", MSDA-Net also has SOTA performance.

  4. !!! We strongly recommend using our integrated framework [PAL Framework], which incorporates multiple networks, including MSDA-Net, and allows convenient switching between different supervision paradigms (Full supervision and Single point supervision) as well as multiple datasets (SIRST, NUDT-SIRST, IRSTD-1k and SIRST3). In addition, compared with this released codebase that resizes training images to a fixed resolution, the PAL framework adopts a cropping strategy with a patch size of 256×256 pixels. Extensive empirical evaluations demonstrate that the PAL framework provides a more reasonable basic training framework and consistently achieves superior performance.

Overview

Overall structure of MSDA-Net
MSDA-Net

We propose a multi-scale direction-aware network (MSDA-Net), which is the first attempt to integrate the high-frequency directional features of infrared small targets as domain prior knowledge into neural networks.
  • High-frequency direction injection module:

High-frequency direction injection module
High-frequency direction injection module

  • Multi-directional feature awareness module:

Multi-directional feature awareness module
Multi-directional feature awareness module

  • Multi-Scale Direction-Aware module:

Multi-Scale Direction-Aware module
Multi-Scale Direction-Aware module

  • Multi-Scale Direction-Aware module:

The feature alignment fusion module
Feature alignment fusion module

-->

Datasets

  1. Original datasets
  1. The datasets we created from original datasets (can be used directly in our demo)

How to use our code

  1. Download the dataset

        Click download datasets

        Unzip the downloaded compressed package to the root directory of the project.

  1. Creat a Anaconda Virtual Environment

    conda create -n MSDA-Net python=3.8 
activate MSDA-Net

  2. Configure the running environment

     pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116
 pip install segmentation_models_pytorch -i https://pypi.tuna.tsinghua.edu.cn/simple
 pip install PyWavelets -i https://pypi.tuna.tsinghua.edu.cn/simple
 pip install scikit-image -i https://pypi.tuna.tsinghua.edu.cn/simple
 pip install albumentations==1.3.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
 pip install scikit-learn matplotlib thop h5py SimpleITK scikit-image medpy yacs torchinfo

  3. Training the model

    The default dataset is IRSTD-1K. If you need to train others, please modify the corresponding dataset path.
    (If you want to train NUDT-SIRST, please note that the parameter IMAGE_SIZE in the training file needs to be changed to 256.)

    python train_MSDA_Net.py

  4. Testing the Model

    Click download weight files in the paper (Unzip the downloaded package to the root directory of the project)

    The default dataset is IRSTD-1K. If you need to test the weight file under other datasets, please modify the corresponding dataset path and weight file path.
    (If you want to train NUDT-SIRST, please note that the parameter IMAGE_SIZE in the training file needs to be changed to 256.)

    python test_MSDA_Net.py

Results and Weights File

  • Quantative Results on NUDT-SIRST:

Comparison of MSDA-Net and various SOTA methods on the NUDT-SIRST dataset
Quantitative Results on NUDT-SIRST

  • Quantative Results on SIRST:

Comparison of MSDA-Net and various SOTA methods on the SIRST dataset
Quantitative Results on SIRST

  • Quantative Results on IRSTD-1k:

Comparison of MSDA-Net and various SOTA methods on the IRSTD-1k dataset
Quantitative Results on IRSTD-1k

Citation

If you find this repo helpful, please give us a 🤩star🤩. Please consider citing the MSDA-Net if it benefits your project.

BibTeX reference is as follows.

@misc{zhao2024multiscaledirectionawarenetworkinfrared,
      title={Multi-Scale Direction-Aware Network for Infrared Small Target Detection}, 
      author={Jinmiao Zhao and Zelin Shi and Chuang Yu and Yunpeng Liu},
      year={2024},
      eprint={2406.02037},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2406.02037}, 
}

word reference is as follows.

Jinmiao Zhao, Zelin Shi, Chuang Yu, and Yunpeng Liu. Multi-Scale Direction-Aware Network for Infrared Small Target Detection. arXiv preprint arXiv:2406.02037, 2024.
  1. My homepage: [YuChuang]
  2. My "PAL Framework" project code (ICCV2025)：[Link]
  3. My latest "FDEP Framework" project code: [Link]