Scholarlink

<p>Cross-view geo-localization between drone and satellite imagery has long been challenged by severe geometric distortion and noise. This paper proposes AFGL-Net, a robust framework that integrates spectral analysis and adversarial learning. It comprises three co-designed components: the Real-Symmetric Frequency Module (RSFM) for efficient feature refinement, the Adversarial Distractor Mining Module (ADMM) for active background suppression, and the Dual-path Robust Cross-Entropy (DRCE) loss for robust optimization. These components collectively learn highly discriminative, view-invariant features. Extensive experiments on the University-1652 and SUES-200 benchmarks establish new state-of-the-art performance, achieving a Recall@1 of 95.86% for satellite-to-drone retrieval. The framework’s design, centered on parallelizable spectral transforms and adversarial mechanisms, is inherently suitable for High-Performance Computing (HPC) and distributed system acceleration, addressing the critical need for high-throughput, real-time processing in large-scale surveillance and emergency response scenarios. The code and model files are publicly available at https://github.com/heliop710823-gif/AFGL-Net.</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

AFGL-Net: adversarial frequency geo-localization network for robust cross-view geo-localization alignment

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • He Xiao,
  • Guang Yang,
  • Qiuming Liu

摘要

Cross-view geo-localization between drone and satellite imagery has long been challenged by severe geometric distortion and noise. This paper proposes AFGL-Net, a robust framework that integrates spectral analysis and adversarial learning. It comprises three co-designed components: the Real-Symmetric Frequency Module (RSFM) for efficient feature refinement, the Adversarial Distractor Mining Module (ADMM) for active background suppression, and the Dual-path Robust Cross-Entropy (DRCE) loss for robust optimization. These components collectively learn highly discriminative, view-invariant features. Extensive experiments on the University-1652 and SUES-200 benchmarks establish new state-of-the-art performance, achieving a Recall@1 of 95.86% for satellite-to-drone retrieval. The framework’s design, centered on parallelizable spectral transforms and adversarial mechanisms, is inherently suitable for High-Performance Computing (HPC) and distributed system acceleration, addressing the critical need for high-throughput, real-time processing in large-scale surveillance and emergency response scenarios. The code and model files are publicly available at https://github.com/heliop710823-gif/AFGL-Net.