Scholarlink

<p>Accurate detection of small objects in unmanned aerial vehicle (UAV) imagery remains a significant challenge in computer vision, primarily due to limited pixel coverage and weak feature representation for targets occupying fewer than 50 pixels. Current state-of-the-art methods face critical limitations: conventional detectors like YOLO rely on static feature concatenation failing to adaptively integrate multi-scale information; popular attention mechanisms (CBAM, SENet) depend on global pooling operations that lose fine-grained spatial details essential for precise localization; and there exists a notable trade-off between accuracy and computational efficiency for resource-constrained UAV platforms. To address these issues, this paper introduces BDMA-YOLO, a parameter-efficient framework for edge deployment that incorporates two new components. The Bidirectional Lightweight Fusion Network (BLFNet) replaces standard connections with dynamic weighting using swish activation, achieving adaptive feature fusion across scales. The Partial Attention Cross-Stage Fusion Block (PACSFB) combines the Multi-Scale Residual Mobile Block (MRMB) with directional spatial attention to preserve key local features. Experiments on the SIMD dataset demonstrate that BDMA-YOLO achieves notable improvements over YOLOv8-S baseline: mAP@0.5 increases by 2.1–81.9%, mAP@0.5:0.95 improves by 2.1–67.0%, and recall rises by 3.9–81.5%. These gains are attained with a 7.2% reduction in parameters and only 1.4% increase in computational load, demonstrating competitive performance compared to recent models such as YOLOv10-S and RT-DETR-L in efficiency–accuracy balance. The source code is available at: <a href="https://github.com/fhxf-pro/BDMA-YOLO">https://github.com/fhxf-pro/BDMA-YOLO</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

Enhancing UAV aerial small target detection with bidirectional dynamic multi-scale attention YOLO

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月27日

查看全文
  • Zhenzhu Wang,
  • Jinzhao Lin,
  • Yu Han,
  • Yu Pang

摘要

Accurate detection of small objects in unmanned aerial vehicle (UAV) imagery remains a significant challenge in computer vision, primarily due to limited pixel coverage and weak feature representation for targets occupying fewer than 50 pixels. Current state-of-the-art methods face critical limitations: conventional detectors like YOLO rely on static feature concatenation failing to adaptively integrate multi-scale information; popular attention mechanisms (CBAM, SENet) depend on global pooling operations that lose fine-grained spatial details essential for precise localization; and there exists a notable trade-off between accuracy and computational efficiency for resource-constrained UAV platforms. To address these issues, this paper introduces BDMA-YOLO, a parameter-efficient framework for edge deployment that incorporates two new components. The Bidirectional Lightweight Fusion Network (BLFNet) replaces standard connections with dynamic weighting using swish activation, achieving adaptive feature fusion across scales. The Partial Attention Cross-Stage Fusion Block (PACSFB) combines the Multi-Scale Residual Mobile Block (MRMB) with directional spatial attention to preserve key local features. Experiments on the SIMD dataset demonstrate that BDMA-YOLO achieves notable improvements over YOLOv8-S baseline: mAP@0.5 increases by 2.1–81.9%, mAP@0.5:0.95 improves by 2.1–67.0%, and recall rises by 3.9–81.5%. These gains are attained with a 7.2% reduction in parameters and only 1.4% increase in computational load, demonstrating competitive performance compared to recent models such as YOLOv10-S and RT-DETR-L in efficiency–accuracy balance. The source code is available at: https://github.com/fhxf-pro/BDMA-YOLO