Scholarlink

<p>The escalating frequency of natural disasters has led to a surge in water-related accidents, which leads to a rapid and accurate detection requires of trapped individuals in water disaster. Addressing the challenges posed by small objects, variable postures and environmental interference, this paper proposed H2Net, a novel enhanced object detection model based on the RT-DETR. It incorporates orthogonal channel attention (OCA) mechanism to enhance feature independence and suppress interference from water surface reflections and waves. Additionally, a wavelet domain fusion module is designed to logarithmically expand the receptive field, improving the detection ability of small objects. Experimental results on the SeaDronesSee-V2 dataset demonstrate that H2Net achieves mAP50, mAP50:95, recall and precision of 85.2, 52.0, 82.1, and 92.1%, respectively, outperforming the baseline. Furthermore, H2Net reduces parameters and computational cost by 22.7 and 6.6%, respectively. And the frame rate is 83 FPS. Cross-domain experiments also validate the effectiveness and generalization ability of H2Net. The results provide efficient technical support for real-time object detection in water disaster scenarios. Our source code is available at: <a href="https://github.com/wangfan040803/H2Net">https://github.com/wangfan040803/H2Net</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

H2Net: enhanced object detection in water disaster scenarios via orthogonal channel attention and wavelet fusion

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Wenzhuo Chen,
  • Wantao Dai,
  • Fan Wang,
  • Long Jin,
  • Peiying Zhao

摘要

The escalating frequency of natural disasters has led to a surge in water-related accidents, which leads to a rapid and accurate detection requires of trapped individuals in water disaster. Addressing the challenges posed by small objects, variable postures and environmental interference, this paper proposed H2Net, a novel enhanced object detection model based on the RT-DETR. It incorporates orthogonal channel attention (OCA) mechanism to enhance feature independence and suppress interference from water surface reflections and waves. Additionally, a wavelet domain fusion module is designed to logarithmically expand the receptive field, improving the detection ability of small objects. Experimental results on the SeaDronesSee-V2 dataset demonstrate that H2Net achieves mAP50, mAP50:95, recall and precision of 85.2, 52.0, 82.1, and 92.1%, respectively, outperforming the baseline. Furthermore, H2Net reduces parameters and computational cost by 22.7 and 6.6%, respectively. And the frame rate is 83 FPS. Cross-domain experiments also validate the effectiveness and generalization ability of H2Net. The results provide efficient technical support for real-time object detection in water disaster scenarios. Our source code is available at: https://github.com/wangfan040803/H2Net.