[IEEE Trans. TGRS 2025] SDS-Net: Shallow-Deep Synergism-detection Network for Infrared Small Target Detection

Overview

SDS-Net is a deep learning framework designed for infrared small target detection (IRSTD). The network employs a shallow-deep synergism strategy that effectively combines multi-scale features from different depths of the network to enhance detection performance.

Main Features

• Shallow-Deep Synergism Architecture: Integrates features from both shallow and deep layers for better target-background discrimination
• Multi-Scale Feature Fusion: Utilizes MSCA (Multi-Scale Cross Attention) and MSSA (Multi-Scale Self Attention) modules
• Deep Supervision: Implements deep supervision during training for better convergence
• Comprehensive Evaluation: Supports multiple evaluation metrics including mIoU, nIoU, Pd, Fa, and F1-score

Architecture

The SDS-Net architecture consists of:

1. Backbone Network: Feature extraction backbone with ResNet-style blocks
2. Shallow Module: Processes early-stage features with cross-attention mechanisms
3. Deep Module: Handles high-level semantic features with self-attention
4. Fusion Blocks: Combines features from different depths using ADSF (Adaptive Deep-Shallow Fusion)
5. Multi-Scale Attention Mechanisms: MSCA and MSSA for effective feature interaction

Results

Evaluation Metrics

The framework provides comprehensive evaluation metrics:

• mIoU: Mean Intersection over Union
• nIoU: Normalized IoU
• Pd: Probability of Detection
• Fa: False Alarm Rate
• F1-score: Harmonic mean of precision and recall

Performance of efficiency and Memory consuming

Supported Datasets

The framework supports multiple IRSTD datasets:

• NUAA-SIRST [download]
• NUDT-SIRST [download]
• IRSTD-1K [download]
• NUDT-SIRST-Sea (for maritime target detection)
• IRDST-real (real-world infrared dataset)

Usage

1. Environment Setup

Option A: Using pip with pyproject.toml (Recommended)

# Clone the repository
git clone https://github.com/PhysiLearn/SDS-Net.git
cd SDS-Net

# Create virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install the package with dependencies
pip install .

# Or install in editable mode for development
pip install -e .

Option B: Using conda (Alternative)

# Create conda environment
conda create -n sdsnet python=3.8
conda activate sdsnet

# Install PyTorch (adjust for your CUDA version)
conda install pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia

# Install other dependencies
pip install einops ml_collections thop
pip install scikit-image Pillow opencv-python
pip install tensorboard tqdm matplotlib

Option C: Manual Installation

# Install core dependencies
pip install torch torchvision
pip install einops ml_collections thop
pip install scikit-image Pillow opencv-python
pip install tensorboard tqdm matplotlib

2. Data Preparation

Organize your dataset in the following structure:

datasets/
├── NUAA-SIRST/
│   ├── images/
│   ├── masks/
│   └── img_idx/
│       ├── train_NUAA-SIRST.txt
│       └── test_NUAA-SIRST.txt
├── NUDT-SIRST/
│   ├── images/
│   ├── masks/
│   └── img_idx/
│       ├── train_NUDT-SIRST.txt
│       └── test_NUDT-SIRST.txt
└── IRSTD-1K/
    ├── images/
    ├── masks/
    └── img_idx/
        ├── train_IRSTD-1K.txt
        └── test_IRSTD-1K.txt

3. Training

Using the command-line entry point (after pip install):

sds-train --dataset_names ['NUDT-SIRST'] --model_names ['SDSNet'] --epochs 1000

Using the Python script directly:

python train.py 

Key Arguments:

• --dataset_names: List of datasets to train on
• --model_names: Model architecture (default: ['SDSNet'])
• --batchSize: Training batch size (default: 16)
• --patchSize: Training patch size (default: 256)
• --epochs: Number of training epochs (default: 1000)
• --dataset_dir: Dataset directory path (default: './datasets')
• --save: Checkpoint save path (default: './log')
• --log_dir: TensorBoard log directory (default: './otherlogs/SDSNet')

4. Testing

Using the command-line entry point (after pip install):

sds-test --dataset_names ['NUDT-SIRST'] --model_names ['SDSNet'] --pth_dirs ['NUDT-SIRST/SDSNet_673_best.pth.tar']

Using the Python script directly:

python test.py 

Project Structure

SDS-Net/
├── model/
│   ├── SDSNet.py          # Main model architecture
│   └── Config.py          # Model configuration
├── images/                # Performance charts and visualizations
├── dataset.py             # Dataset loading utilities
├── train.py              # Training script
├── test.py               # Testing and evaluation script
├── metrics.py            # Evaluation metrics
├── utils.py              # Utility functions
├── warmup_scheduler.py   # Learning rate scheduling
├── pyproject.toml        # Modern Python project configuration
└── README.md             # This file

Configuration

The model configuration is defined in model/Config.py:

config = {
    'transformer': {
        'num_heads': 4,
        'num_layers': 4
    },
    'patch_sizes': [16, 8, 4],
    'base_channel': 32,
    'n_classes': 1
}

Citation

If you find this work useful, please consider citing:

@article{SDSNet2025,
  title={SDS-Net: Shallow–Deep Synergism-Detection Network for Infrared Small Target Detection},
  author={Yue, Taoran and Lu, Xiaojin and Cai, Jiaxi and Chen, Yuanping and Chu, Shibing},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  volume={63},
  pages={1–13},
  year={2025},
  publisher={Institute of Electrical and Electronics Engineers (IEEE)},
  ISSN={1558-0644},
  DOI={10.1109/tgrs.2025.3588117},
  url={http://dx.doi.org/10.1109/tgrs.2025.3588117}
}

Paper Title: SDS-Net: Shallow-Deep Synergism-detection Network for Infrared Small Target Detection

Contact

For questions or issues regarding SDS-Net, please contact the development team or open an issue on GitHub.

Acknowledgments

This implementation is inspired by recent advances in infrared small target detection and builds upon the following works:

• IRSTD-Toolbox
• SCTransNet
• UCTransNet
• DNA-Net
• UIU-Net

License

This project is released under the MIT License.