Scholarlink

<p>The characteristics of remote sensing images, such as complex ground objects, rich feature details, large intraclass variance, and small interclass variance, usually require deep learning semantic segmentation methods to have strong feature learning representation ability. While convolutional neural networks (CNNs) excel at local feature extraction through convolutional operations, their limited receptive fields hinder long-range dependency modeling. Transformers address this through multi-head self-attention but introduce prohibitive computational costs. We propose RFCFormer, an encoder–decoder network introducing: (1) rectangular dynamic self-integration module (RDIM) for spatial reconstruction and pyramid context extraction and (2) Fourier convolution former block (FCFB) with Fourier attention mixer (FCAM) to enhance feature representation. Evaluations on Vaihingen/Potsdam benchmarks show RFCFormer achieves 84.64%/87.32% mIoU, outperforming state-of-the-art methods by 0.5%/0.42% points.</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

RFCFormer:rectangular Fourier convolution former for remote sensing semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Jiayin Ding,
  • Wenbin Shi,
  • Jingsheng Lei,
  • Yong Ji

摘要

The characteristics of remote sensing images, such as complex ground objects, rich feature details, large intraclass variance, and small interclass variance, usually require deep learning semantic segmentation methods to have strong feature learning representation ability. While convolutional neural networks (CNNs) excel at local feature extraction through convolutional operations, their limited receptive fields hinder long-range dependency modeling. Transformers address this through multi-head self-attention but introduce prohibitive computational costs. We propose RFCFormer, an encoder–decoder network introducing: (1) rectangular dynamic self-integration module (RDIM) for spatial reconstruction and pyramid context extraction and (2) Fourier convolution former block (FCFB) with Fourier attention mixer (FCAM) to enhance feature representation. Evaluations on Vaihingen/Potsdam benchmarks show RFCFormer achieves 84.64%/87.32% mIoU, outperforming state-of-the-art methods by 0.5%/0.42% points.