Scholarlink

<p>Underwater object detection is hindered by optical effects like light attenuation and color shift, while traditional RGB-D fusion methods struggle to model inter-modal semantic relationships, resulting in poor robustness in complex environments. To address this, we propose MCAB-YOLO, a lightweight dual-modality detection framework. Its core, the Multimodal Cross-Attention Block (MCAB), employs a parallel architecture combining a parameter-free self-enhancement mechanism with a window-based cross-modal attention module to achieve fine-grained feature interaction and adaptive fusion. Additionally, the backbone is reconstructed using depthwise separable convolutions to reduce computational overhead. Experiments show that our model achieves a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textrm{mAP}_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation> of 94.1% on a custom dataset, outperforming YOLO-Concat and YOLOv11 by 1.8% and 2.7%, respectively. Furthermore, it enables millisecond-level inference on low-power GPUs, significantly enhancing detection performance in resource-constrained 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

MCAB-YOLO: multimodal cross-attention for RGB-D underwater object detection

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Haiyong Wang,
  • Zheng Han

摘要

Underwater object detection is hindered by optical effects like light attenuation and color shift, while traditional RGB-D fusion methods struggle to model inter-modal semantic relationships, resulting in poor robustness in complex environments. To address this, we propose MCAB-YOLO, a lightweight dual-modality detection framework. Its core, the Multimodal Cross-Attention Block (MCAB), employs a parallel architecture combining a parameter-free self-enhancement mechanism with a window-based cross-modal attention module to achieve fine-grained feature interaction and adaptive fusion. Additionally, the backbone is reconstructed using depthwise separable convolutions to reduce computational overhead. Experiments show that our model achieves a \(\textrm{mAP}_{50}\) mAP 50 of 94.1% on a custom dataset, outperforming YOLO-Concat and YOLOv11 by 1.8% and 2.7%, respectively. Furthermore, it enables millisecond-level inference on low-power GPUs, significantly enhancing detection performance in resource-constrained scenarios.