Scholarlink

<p>High-resolution remote sensing instance segmentation faces challenges such as scale variations, orientation ambiguities, and background clutter, which hinder accurate object localization and mask prediction. To address these issues, we propose the bidirectional attention fusion network (BiAFNet), which employs a dual-path context fusion paradigm to balance global semantics and local details through spatial-channel attention mechanisms. The dual-path framework establishes bidirectional interactions between primary and secondary paths, addressing scale variations and improving detection of small-scale/dense targets. The dual attention fusion module (DAFM) integrates direction-aware attention to model anisotropic object distributions, explicitly tackling orientation ambiguities and clustered distributions. The frequency-domain enhanced simplified hierarchical region of interest extractor (SHRoIE) suppresses background clutter by decomposing features into structural and noise components in the spectral domain. The integration of these components collaboratively optimizes multi-scale feature generalization, achieving improved performance in complex remote sensing scenarios compared to baseline approaches. Experiments on NWPU VHR-10, DIOR, and HRSID datasets demonstrate that BiAFNet achieves mean Average Precision (mAP) scores of 73.2%, 73.4%, and 57.7%, respectively, validating its effectiveness in complex remote sensing scenarios. Code is available at <a href="https://github.com/Tsing-time/BiAFNet">https://github.com/Tsing-time/BiAFNet</a>.</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

Dual-flow context enhancement for precise instance segmentation in remote sensing imagery

  • 期刊论文

  • 发表日期：

    2025年12月6日

查看全文
  • Dongjie Zhou,
  • Xingxing Yao,
  • Ziyi Yu,
  • Zhaohui Zheng

摘要

High-resolution remote sensing instance segmentation faces challenges such as scale variations, orientation ambiguities, and background clutter, which hinder accurate object localization and mask prediction. To address these issues, we propose the bidirectional attention fusion network (BiAFNet), which employs a dual-path context fusion paradigm to balance global semantics and local details through spatial-channel attention mechanisms. The dual-path framework establishes bidirectional interactions between primary and secondary paths, addressing scale variations and improving detection of small-scale/dense targets. The dual attention fusion module (DAFM) integrates direction-aware attention to model anisotropic object distributions, explicitly tackling orientation ambiguities and clustered distributions. The frequency-domain enhanced simplified hierarchical region of interest extractor (SHRoIE) suppresses background clutter by decomposing features into structural and noise components in the spectral domain. The integration of these components collaboratively optimizes multi-scale feature generalization, achieving improved performance in complex remote sensing scenarios compared to baseline approaches. Experiments on NWPU VHR-10, DIOR, and HRSID datasets demonstrate that BiAFNet achieves mean Average Precision (mAP) scores of 73.2%, 73.4%, and 57.7%, respectively, validating its effectiveness in complex remote sensing scenarios. Code is available at https://github.com/Tsing-time/BiAFNet.