Scholarlink

<p>The current advanced visual model detectors usually adopt Feature Pyramid Networks (FPN) to extract multi-scale information. However, the classic FPN and its variants (such as AUGFPN, PAFPN, etc.) do not consider the spatial misalignment of multi-scale features at the global scale, which results in detectors being unable to locate targets with the best accuracy optimally. This paper proposes a novel Bidirectional Feature Aligned Feature Pyramid Network (BAFPN). The network progressively aligns targets’ spatial location and shape across all levels of features to match the original image through a preceding bottom-up information propagation path, combined with a Spatial Feature Alignment Module (SPAM), thus fundamentally solving the above problem. In the top-down cross-scale feature fusion path, BAFPN reduces the semantic gap between adjacent feature maps while maintaining feature expression diversity through a Fine-Grained Semantic Alignment Module (SEAM), mitigating the aliasing effect caused by cross-scale feature fusion and maximizing the potential of FPN. Additionally, we replace FPN’s 1<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>1 convolutional layer lateral connection module with a Grouped Aggregation Lateral Connection Module (GALM), which reduces the number of feature channels while retaining the rich channel information of the original features output by the backbone. BAFPN significantly improves the accuracy of target localization for various detectors. On the DOTAv1.5 dataset, BAFPN improves the AP75 of the baseline model by 1.68 percentage points and the AP50 and mAP by 1.45 and 1.34 percentage points, respectively.</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

BAFPN: bidirectionally aligning features to improve object localization accuracy in remote sensing images

  • 期刊论文

  • 发表日期：

    2025年11月2日

查看全文
  • Jiakun Li,
  • Qingqing Wang,
  • Hongbin Dong

摘要

The current advanced visual model detectors usually adopt Feature Pyramid Networks (FPN) to extract multi-scale information. However, the classic FPN and its variants (such as AUGFPN, PAFPN, etc.) do not consider the spatial misalignment of multi-scale features at the global scale, which results in detectors being unable to locate targets with the best accuracy optimally. This paper proposes a novel Bidirectional Feature Aligned Feature Pyramid Network (BAFPN). The network progressively aligns targets’ spatial location and shape across all levels of features to match the original image through a preceding bottom-up information propagation path, combined with a Spatial Feature Alignment Module (SPAM), thus fundamentally solving the above problem. In the top-down cross-scale feature fusion path, BAFPN reduces the semantic gap between adjacent feature maps while maintaining feature expression diversity through a Fine-Grained Semantic Alignment Module (SEAM), mitigating the aliasing effect caused by cross-scale feature fusion and maximizing the potential of FPN. Additionally, we replace FPN’s 1 \(\times\) 1 convolutional layer lateral connection module with a Grouped Aggregation Lateral Connection Module (GALM), which reduces the number of feature channels while retaining the rich channel information of the original features output by the backbone. BAFPN significantly improves the accuracy of target localization for various detectors. On the DOTAv1.5 dataset, BAFPN improves the AP75 of the baseline model by 1.68 percentage points and the AP50 and mAP by 1.45 and 1.34 percentage points, respectively.