Tested on Windows 10 , with Python 3.7, PyTorch 1.13, NVIDIA 3080.
The infrared small target public dataset:
The IRSTD-1k dataset
The MDvsFA dataset
The IRST640 dataset
The bounding box annotation version of the current infrared small target public dataset: download from BaiduYun Drive with code IRST or Google Drive.
pip install -r requirements.txt
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"
train:
Download the dataset and put it in the data file
python train.py --workers 0 --device 0 --batch-size 8 --data data/NUAA-sirst.yaml --img 640 640 --cfg cfg/EFL.yaml --weights '' --name NUAA --hyp data/hyp.scratch.p5.yaml
test:
python test.py --data data/NUAA-sirst.yaml --img 640 --batch 32 --conf 0.001 --iou 0.5 --device 0 --weights NUAA.pt --name NUAA
inference:
python detect.py --weights runs/train/NUAA.pt --conf 0.5 --img-size 640 --source data/NUAA-sirst/images/test
Method |
NUAA-SIRST |
NUDT-SIRST |
IRSTD-1k |
||||||
Pre |
Rec |
F1 |
Pre |
Rec |
F1 |
Pre |
Rec |
F1 |
|
MDvsFA |
0.845 |
0.507 |
0.597 |
0.608 |
0.192 |
0.262 |
0.55 |
0.483 |
0.475 |
AGPCNet |
0.39 |
0.81 |
0.527 |
0.368 |
0.684 |
0.479 |
0.415 |
0.47 |
0.441 |
ACM |
0.765 |
0.762 |
0.763 |
0.732 |
0.745 |
0.738 |
0.679 |
0.605 |
0.64 |
ISNet |
0.82 |
0.847 |
0.834 |
0.742 |
0.834 |
0.785 |
0.718 |
0.741 |
0.729 |
ACLNet |
0.848 |
0.78 |
0.813 |
0.868 |
0.772 |
0.817 |
0.843 |
0.656 |
0.738 |
DNANet |
0.847 |
0.836 |
0.841 |
0.914 |
0.889 |
0.901 |
0.768 |
0.721 |
0.744 |
ours |
0.882 |
0.858 |
0.870 |
0.963 |
0.931 |
0.947 |
0.870 |
0.817 |
0.843 |
Download |
@article{yang2024eflnet,
title={EFLNet: Enhancing Feature Learning Network for Infrared Small Target Detection},
author={Yang, Bo and Zhang, Xinyu and Zhang, Jian and Luo, Jun and Zhou, Mingliang and Pi, Yangjun},
journal={IEEE Transactions on Geoscience and Remote Sensing},
volume={62},
pages={1--11},
year={2024},
publisher={IEEE}
}