Scholarlink

<p>This paper presents DBA-YOLO, a dual-branch axial feature enhancement network based on YOLOv12, aimed at small object detection in complex traffic scenarios. The model addresses misdetection and missed detection of small objects due to lighting variations and occlusions. Key contributions include: 1) the MLAA (Mamba-Like Axial Attention) module, which integrates a Mamba-style gating mechanism and achieves accurate small object detection through spatial-channel and segmented axial attention; 2) the HLGAB (Hierarchical Local-Global Attention Block) module, which enhances feature interaction via dual-path channels, reducing misdetection in dense scenes. Experiments on the BDD100K dataset show a <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4572_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="55" /> </InlineMediaObject> <EquationSource Format="TEX">\( mAP_{0.5} \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>m</mi> <mi>A</mi> <msub> <mi>P</mi> <mrow> <mn>0.5</mn> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> of 48.43%, a 1.87% improvement over YOLOv12n, with real-time detection performance. Visual analysis demonstrates robustness in challenging conditions like nighttime and fog. Additionally, a supporting system is developed for real-time detection using image stream processing. The code and dataset will be available at <a href="https://github.com/ZehuaChenLab/DBA-YOLO">https://github.com/ZehuaChenLab/DBA-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

DBA-YOLO: a mamba-incorporated dual-branch axial feature enhancement network for small object detection in traffic scenarios

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Jing Zhang,
  • Chenxin Zhang,
  • Lianghao Xu,
  • BeiBei Mao,
  • Zehua Chen

摘要

This paper presents DBA-YOLO, a dual-branch axial feature enhancement network based on YOLOv12, aimed at small object detection in complex traffic scenarios. The model addresses misdetection and missed detection of small objects due to lighting variations and occlusions. Key contributions include: 1) the MLAA (Mamba-Like Axial Attention) module, which integrates a Mamba-style gating mechanism and achieves accurate small object detection through spatial-channel and segmented axial attention; 2) the HLGAB (Hierarchical Local-Global Attention Block) module, which enhances feature interaction via dual-path channels, reducing misdetection in dense scenes. Experiments on the BDD100K dataset show a \( mAP_{0.5} \) m A P 0.5 of 48.43%, a 1.87% improvement over YOLOv12n, with real-time detection performance. Visual analysis demonstrates robustness in challenging conditions like nighttime and fog. Additionally, a supporting system is developed for real-time detection using image stream processing. The code and dataset will be available at https://github.com/ZehuaChenLab/DBA-YOLO.