Scholarlink

<p>Marine debris is a major component of the global environmental problem, seriously affecting aquatic life’s survival and ecosystems’ health. They are difficult to degrade and migrate over long distances, and most of them enter the deep sea and remain there. Autonomous underwater robots can remove deep-sea debris to some extent. To detect deep-sea debris accurately and quickly, a lightweight detection network (Composite Deep-sea Debris YOLO, CDsd-YOLO) based on You Only Look Once v8n (YOLOv8n) is proposed in this paper. In this study, a modified Bidirectional Pyramid Network (BiPN) is used to enhance the model’s ability to discriminate background differences. Then, Efficient-Head is designed to facilitate the deployment of the model into an autonomous underwater robot. Finally, the Wise Distance IoU (Wise-DIoU) loss function is proposed to reduce the impact of low-quality samples on the detection performance. The images in the Composite Deep-sea Debris Dataset (CDsd) are captured by submersibles in real deep-sea environments and have a high degree of realism and have a high degree of fidelity. Experimental results on the CDsd dataset show that the improved model achieves 85.7% and 66.4% average accuracy values at the 0.5 threshold (mAP0.5) and 0.5:0.95 (mAP0.5:0.95) thresholds, respectively, and the number of parameters is reduced by 48.2% compared to YOLOv8n. The mAP0.5 and mAP0.5: 0.95 values were improved by 1.9% and 3.5%, respectively.</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

CDsd-YOLO: a real-time detection method of deep-sea debris based on YOLO

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Biao Zhang,
  • Zhenyang Zhu,
  • Jiazhong Xu

摘要

Marine debris is a major component of the global environmental problem, seriously affecting aquatic life’s survival and ecosystems’ health. They are difficult to degrade and migrate over long distances, and most of them enter the deep sea and remain there. Autonomous underwater robots can remove deep-sea debris to some extent. To detect deep-sea debris accurately and quickly, a lightweight detection network (Composite Deep-sea Debris YOLO, CDsd-YOLO) based on You Only Look Once v8n (YOLOv8n) is proposed in this paper. In this study, a modified Bidirectional Pyramid Network (BiPN) is used to enhance the model’s ability to discriminate background differences. Then, Efficient-Head is designed to facilitate the deployment of the model into an autonomous underwater robot. Finally, the Wise Distance IoU (Wise-DIoU) loss function is proposed to reduce the impact of low-quality samples on the detection performance. The images in the Composite Deep-sea Debris Dataset (CDsd) are captured by submersibles in real deep-sea environments and have a high degree of realism and have a high degree of fidelity. Experimental results on the CDsd dataset show that the improved model achieves 85.7% and 66.4% average accuracy values at the 0.5 threshold (mAP0.5) and 0.5:0.95 (mAP0.5:0.95) thresholds, respectively, and the number of parameters is reduced by 48.2% compared to YOLOv8n. The mAP0.5 and mAP0.5: 0.95 values were improved by 1.9% and 3.5%, respectively.