.png){kind=icon}
.png){kind=icon}
Return
Label noise learning based SAR target classification method
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2025.108373.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The recognition of Synthetic Aperture Radar (SAR) Target is a critical task in SAR image interpretation. With their exceptional capacity to model complex data structures, Convolutional Neural Networks(CNNs) are now the standard architecture for addressing SAR image classification problems. However, these methods typically require large-scale labeled datasets for training. SAR images are inherently susceptible to both feature and label noise due to the technical sophistication of the imaging process and the high likelihood of human error during annotation. This often leads to a significant degradation in the performance of CNN-based classifiers. To mitigate feature noise, we propose a dynamic Lp-norm regularization-based scattering feature extraction method that leverages neural networks to automatically estimate and adapt the regularization parameters at each layer. To address label noise, we further develop a robust representation learning framework for SAR target classification, which enhances model robustness by minimizing the distances between samples and their corresponding class prototypes. Extensive experiments conducted on three widely-used SAR datasets — MSTAR, SAR-ACD, and FUSAR — show that the proposed method consistently achieves robust classification accuracy across label noise levels from 0 % to 60 %, significantly mitigating the adverse effects of annotation inaccuracies.
Journal
IF:
6.3
Papers:
7.8K
Citations:
3.0W