太赫兹科学与电子信息学报2025，Vol.23Issue(11)：1150-1156,7.DOI:10.11805/TKYDA2024500
基于光学像增广的空间目标ISAR像识别方法
Recognition of space target ISAR image based on optical image data augmentation
黄鹤鸣 1杜亮 1王天云 2金林 3程强1
作者信息
- 1. 南京电子技术研究所,江苏 南京 210039||雷达探测感知全国重点实验室,江苏 南京 210039||江苏省探测感知技术重点实验室,江苏 南京 210039
- 2. 中国卫星海上测控部,江苏 无锡 214431
- 3. 南京电子技术研究所,江苏 南京 210039
- 折叠
摘要
Abstract
Aiming at the problem that the ISAR image recognition method of space targets under traditional supervised learning is limited by the number of samples and the quality of annotations,and the model performance cannot meet the expectations,a pre-training-fine-tuning based ISAR image recognition method for space targets is proposed.Firstly,the gap between the optical image source domain and the ISAR image target domain is narrowed by data augmentation means such as grid masking,morphological closing operations,and binarization preprocessing.Subsequently,the general visual backbone pre-trained network ResNeXt-50 is employed to perform supervised learning on the optical images of space targets and their augmented images.The results show that compared with the traditional method,the recognition accuracy on the test set is increased from 21.11%to 72.83%,and the macro-F1 score is increased from 0.124 2 to 0.720 3.The method in this paper significantly improves the model's recognition ability on ISAR images when the number of ISAR samples is small or even lacks annotations.关键词
空间目标/逆合成孔径雷达/迁移学习/领域自适应/预训练模型/数据增强Key words
space target/Inverse Synthetic Aperture Radar(ISAR)/transfer learning/domain adaptation/pre-trained model/data augmentation分类
信息技术与安全科学引用本文复制引用
黄鹤鸣,杜亮,王天云,金林,程强..基于光学像增广的空间目标ISAR像识别方法[J].太赫兹科学与电子信息学报,2025,23(11):1150-1156,7.
