Scholarlink

<p>Accurate object detection in Synthetic Aperture Radar (SAR) imagery remains challenging due to strong speckle, cluttered backgrounds, small targets, and arbitrary orientations. Existing methods usually tackle these factors in isolation—via multi-scale fusion, rotation-robust design, or attention-based background suppression—so their performance often degrades when multiple difficulties co-occur. In this paper, we propose a Global–Dual Attention Network (GDAN) that jointly models multi-scale structure, global context, and local adaptivity within a single detection framework. A Spatial Pyramid Pooling module first builds compact multi-scale features suited to both tiny and large objects. On top of this representation, a Global Attention Mechanism (GAM) aggregates cross-dimensional dependencies to enhance globally consistent semantics, while a Dual Attention (DA) module performs global feature gathering and local redistribution using second-order attention pooling, selectively strengthening target regions and suppressing clutter. Extensive experiments on the HighSAR dataset and the PASCAL VOC benchmark demonstrate that GDAN achieves state-of-the-art performance, improving mAP@0.5 by up to 2.6 percentage points over the strongest competing detector, with consistent gains in both precision and recall. These results confirm that coupling global–dual attention with multi-scale fusion provides a principled and effective solution for robust SAR object detection and generalizes well to natural image scenarios.</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

GDAN: global–dual attention with multi-scale fusion for robust SAR object detection

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Shengjie Lei,
  • Zhiyong Wei,
  • Yulian Zhang,
  • Meihua Fang,
  • Baowen Wu

摘要

Accurate object detection in Synthetic Aperture Radar (SAR) imagery remains challenging due to strong speckle, cluttered backgrounds, small targets, and arbitrary orientations. Existing methods usually tackle these factors in isolation—via multi-scale fusion, rotation-robust design, or attention-based background suppression—so their performance often degrades when multiple difficulties co-occur. In this paper, we propose a Global–Dual Attention Network (GDAN) that jointly models multi-scale structure, global context, and local adaptivity within a single detection framework. A Spatial Pyramid Pooling module first builds compact multi-scale features suited to both tiny and large objects. On top of this representation, a Global Attention Mechanism (GAM) aggregates cross-dimensional dependencies to enhance globally consistent semantics, while a Dual Attention (DA) module performs global feature gathering and local redistribution using second-order attention pooling, selectively strengthening target regions and suppressing clutter. Extensive experiments on the HighSAR dataset and the PASCAL VOC benchmark demonstrate that GDAN achieves state-of-the-art performance, improving mAP@0.5 by up to 2.6 percentage points over the strongest competing detector, with consistent gains in both precision and recall. These results confirm that coupling global–dual attention with multi-scale fusion provides a principled and effective solution for robust SAR object detection and generalizes well to natural image scenarios.