A Prototype-Based Semisupervised Learning Method for Few-Shot SAR Target Recognition

doi:10.1109/JSTARS.2025.3643525

A Prototype-Based Semisupervised Learning Method for Few-Shot SAR Target Recognition

2026-01-01

0

OA

AI

DOI：10.1109/JSTARS.2025.3643525

Original

Save

View PDF

Abstract

En 中文

Deep learning-based methods have achieved extraordinary success in SAR automatic target recognition. However, deep learning conventionally necessitates a substantial number of labeled samples to achieve effective training, and labeled samples of new classes in real-world scenarios are scarce, which limits the performance of existing methods in the few-shot task. In response to this issue, this article proposes a prototype-based semisupervised learning method for few-shot SAR target recognition, named WST-DRFSL. The method consists of two stages: the base learning stage and the dynamic refinement (DR) stage. In the first stage, a robust encoder is trained on both labeled and unlabeled samples of base classes via consistency regularization. Then, in the second stage, pseudolabels and CR are iteratively applied to new classes' few labeled samples and abundant unlabeled samples to achieve superior new-class recognition performance. Furthermore, the wavelet scattering transform (WST) is employed in both stages to fully exploit the scattering characteristics of SAR images. Extensive simulations on MSTAR, FUSAR, OpenSARShip, and SAMPLE datasets have demonstrated that the proposed method surpasses the state-of-the-art recognition accuracy on the few-shot learning tasks. The code is available online.

Keywords:

Few-shot learning (FSL)

synthetic aperture radar (SAR) automatic target recognition

semisupervised learning

A Prototype-Based Semisupervised Learning Method for Few-Shot SAR Target Recognition

Abstract

AI Summary

Journal

Organization