Scholarlink

<p>This paper proposes CSWin-WNet, a W-shaped deep network based on the CSWin Transformer, for noise-robust inverse synthetic aperture radar (ISAR) imaging under challenging conditions. The method starts with an initial reconstruction using the 2D-SL0 algorithm, followed by a Bayesian denoising framework that enables the derivation of posteriors and the variational lower bound. Inspired by the strengths of U-Net and the CSWin Transformer, CSWin-WNet is designed to jointly estimate the posterior distribution of the reconstructed image and the noise variance. By integrating sparse Bayesian learning with deep learning, the negative variational lower bound is used as the loss function, allowing end-to-end training on simulated data with varying noise levels to obtain optimal network parameters. Experimental results on the proprietary aircraft ISAR dataset (AID) and satellite ISAR dataset (SID) show that CSWin-WNet achieves superior imaging performance, enhanced noise robustness, and strong generalization capability. These results confirm its effectiveness and reliability for ISAR imaging applications under low signal-to-noise ratio and incomplete data conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CSWin-WNet: a W-shaped network based on CSWin Transformer for noise-robust ISAR imaging

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Xuemei Ren,
  • Xiaoyong Li,
  • Zhenbiao Zhang,
  • Lei Liu,
  • Xueru Bai,
  • Feng Zhou

摘要

This paper proposes CSWin-WNet, a W-shaped deep network based on the CSWin Transformer, for noise-robust inverse synthetic aperture radar (ISAR) imaging under challenging conditions. The method starts with an initial reconstruction using the 2D-SL0 algorithm, followed by a Bayesian denoising framework that enables the derivation of posteriors and the variational lower bound. Inspired by the strengths of U-Net and the CSWin Transformer, CSWin-WNet is designed to jointly estimate the posterior distribution of the reconstructed image and the noise variance. By integrating sparse Bayesian learning with deep learning, the negative variational lower bound is used as the loss function, allowing end-to-end training on simulated data with varying noise levels to obtain optimal network parameters. Experimental results on the proprietary aircraft ISAR dataset (AID) and satellite ISAR dataset (SID) show that CSWin-WNet achieves superior imaging performance, enhanced noise robustness, and strong generalization capability. These results confirm its effectiveness and reliability for ISAR imaging applications under low signal-to-noise ratio and incomplete data conditions.