Scholarlink

<p>Multi-scale feature fusion, as a key component in remote-sensing rotated object detection, often relies on CNN and Transformer for feature fusion. However, in the fusion stage, CNN-based methods focus on detail perception but lack global perception capability, while Transformer-based methods emphasize global perception but lack detail perception capability. In this paper, we propose a multi-scale feature fusion method based on the biological visual attention mechanism&#xa0;(BVAMFPN). By simultaneously preserving local detailed features and global feature information, it enables the model to perceive densely packed small instances more clearly. First, feature maps from different layers are input into the Biological Visual Attention Mechanism&#xa0;(BVAM) following a deep-shallow interactive fusion order. Second, to ensure consistent spatial dimensions among features from various layers, bilinear interpolation resamples the input. Finally, a biologically inspired visual attention mechanism employs a sliding window approach to extract local features from cross-layer feature maps, concurrently captures global information via a pooling pathway, and subsequently integrates these cross-layer features with those of the current layer. Multiple models that incorporate BVAMFPN achieve varying degrees of improvement in recognition accuracy. It reduces 4.6 Params (MB) and 46.0 FLOPs (G) compared to the original. Combined with ROI-Transformer, the model achieves a prediction accuracy of 80.4% on DOTA-v1.0 and 76.8% on DOTA-v1.5. The codes are available at <a href="https://github.com/ssdle/BVAMFPN.git">https://github.com/ssdle/BVAMFPN.git</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

BVAMFPN: multi-scale feature fusion for rotated object detection in remote sensing based on biological visual attention mechanism

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Zhiyou Wang,
  • Jingmin Yang,
  • Yongchao Qiao,
  • Wenjie Zhang

摘要

Multi-scale feature fusion, as a key component in remote-sensing rotated object detection, often relies on CNN and Transformer for feature fusion. However, in the fusion stage, CNN-based methods focus on detail perception but lack global perception capability, while Transformer-based methods emphasize global perception but lack detail perception capability. In this paper, we propose a multi-scale feature fusion method based on the biological visual attention mechanism (BVAMFPN). By simultaneously preserving local detailed features and global feature information, it enables the model to perceive densely packed small instances more clearly. First, feature maps from different layers are input into the Biological Visual Attention Mechanism (BVAM) following a deep-shallow interactive fusion order. Second, to ensure consistent spatial dimensions among features from various layers, bilinear interpolation resamples the input. Finally, a biologically inspired visual attention mechanism employs a sliding window approach to extract local features from cross-layer feature maps, concurrently captures global information via a pooling pathway, and subsequently integrates these cross-layer features with those of the current layer. Multiple models that incorporate BVAMFPN achieve varying degrees of improvement in recognition accuracy. It reduces 4.6 Params (MB) and 46.0 FLOPs (G) compared to the original. Combined with ROI-Transformer, the model achieves a prediction accuracy of 80.4% on DOTA-v1.0 and 76.8% on DOTA-v1.5. The codes are available at https://github.com/ssdle/BVAMFPN.git.