HDNet: A Hybrid Domain Network with Multi-Scale High-Frequency Information Enhancement for Infrared Small Target Detection

Official PyTorch implementation of the TGRS 2025 paper "HDNet: A Hybrid Domain Network with Multi-Scale High-Frequency Information Enhancement for Infrared Small Target Detection".

Authors

Mingzhu Xu¹, Chenglong Yu¹, Zexuan Li¹, Haoyu Tang¹, Yupeng Hu¹*, Liqiang Nie²

¹ Shandong University
² Harbin Institute of Technology (Shen Zhen)
* Corresponding author

Links

• Paper: IEEE Xplore
• Code Repository: GitHub

---

Table of Contents

• Introduction
• Highlights
• Method / Framework
• Project Structure
• Installation
• Dataset
• Usage
• Quantitative Results
• Citation
• Acknowledgement
• License

---

Introduction

This project is the official implementation of the paper "HDNet: A Hybrid Domain Network with Multi-Scale High-Frequency Information Enhancement for Infrared Small Target Detection".

HDNet proposes an innovative hybrid-domain network that significantly enhances the contrast of infrared small targets and suppresses background interference by integrating frequency-domain features with traditional spatial-domain CNN features:

• Spatial Branch: Introduces a Multi-scale Atrous Convolution (MAC) module to enhance the perception of small targets at different scales.
• Frequency Branch: Designs a Dynamic High-Pass Filter (DHPF) module to dynamically remove low-frequency background interference while preserving high-frequency target details.
• Experimental Performance: Outperforms 26 state-of-the-art methods on three benchmark datasets: IRSTD-1K, NUAA-SIRST, and NUDT-SIRST.

Example Description

We present HDNet, a framework for Infrared Small Target Detection (IRSTD).
Our method addresses background interference and low contrast by introducing hybrid domain feature fusion and dynamic high-pass filtering.
This repository provides the official implementation, pretrained weights, and evaluation scripts.

---

Highlights

• Proposes a Hybrid-Domain Network (HDNet) that combines spatial multi-scale perception and frequency-domain background suppression.
• Introduces the MAC module, improving contrast between small targets and complex backgrounds.
• Introduces the DHPF module, dynamically suppressing low-frequency background noise.
• Provides complete evaluation results on three public benchmark datasets.

---

Method / Framework

The HDNet architecture illustrates a dual-branch collaboration between spatial and frequency domains.

Framework Figure

Figure 1. Overall framework of HDNet.

---

Project Structure

.
├── Fig/                   # Architecture diagrams and visualization results
├── datasets/              # Datasets (IRSTD-1K, NUAA-SIRST, NUDT-SIRST)
├── weight/                # Pretrained weights (.pkl)
├── main.py                # Main entry script
├── README.md
└── requirements.txt

---

Installation

1. Clone the repository

git clone [https://github.com/iLearn-Lab/HDNet.git](https://github.com/iLearn-Lab/HDNet.git)
cd HDNet

2. Prerequisites

This project was developed under Ubuntu 22.04. Recommended environment:

• Python 3.10
• PyTorch 2.1.0
• CUDA 12.1

pip install -r requirements.txt

---

Dataset

Please download the following datasets and place them in the ./datasets directory:

• IRSTD-1k
• NUAA-SIRST
• NUDT-SIRST

---

Usage

Training

python main.py --dataset-dir './dataset/IRSTD-1k' --batch-size 4 --epochs 800 --mode 'train'

Testing

python main.py --dataset-dir './dataset/IRSTD-1k' --batch-size 4 --mode 'test' --weight-path './weight/irstd.pkl'

---

Quantitative Results

Dataset	mIoU (x10⁻²)	Pd (x10⁻²)	Fa (x10⁻⁶)	Weights
IRSTD-1k	70.26	94.56	4.33	Download
NUAA-SIRST	79.17	100	0.53	Download
NUDT-SIRST	85.17	98.52	2.78	Download

Visualization results can be found here：HDNet_Visual_Result

---

Citation

If you use this code in your research, please cite our paper:

@ARTICLE{11017756,
  author={Xu, Mingzhu and Yu, Chenglong and Li, Zexuan and Tang, Haoyu and Hu, Yupeng and Nie, Liqiang},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={HDNet: A Hybrid Domain Network With Multiscale High-Frequency Information Enhancement for Infrared Small-Target Detection}, 
  year={2025},
  volume={63},
  number={},
  pages={1-15},
  doi={10.1109/TGRS.2025.3574962}
}

---

Acknowledgement

• HDNet adopts the SLS loss function and builds upon the architecture ofMSHNet. Special thanks to the work of Qiankun Liu.

---

License

This project is released under the Apache License 2.0.