Scholarlink

<p>Underwater object recognition is crucial for achieving automated marine fishing and ensuring the sustainable development of marine resources. However, underwater environments’ complexity, such as low visibility and small, densely distributed targets, often causes detection issues like false positives, missed detections, and low accuracy. To tackle these challenges, this study introduces UES-YOLO, an enhanced lightweight underwater target detection algorithm based on YOLOv8n. We developed the Shallow and Maximum Attention Deep Robust Sampling Module (SMAD) to improve feature extraction. SMAD includes Shallow Robust Feature Downsampling (SRFD) for initial image processing and Maximum Attention Deep Robust Feature Downsampling (MADRFD) for deeper feature extraction. A Lightweight Asymmetric Detection Head (LADH) is also added to reduce parameters and computational complexity. The Minimum Point Distance Intersection over Union (MPDIoU) is used to enhance UES-YOLO’s generalization. Tests show UES-YOLO achieves mAP50 scores of 71% and 70.4% on the UDD and Aquarium datasets, improving by 8% and 1.8% over the original model. The mAP50-95 scores are 29.9% and 38.2%, increasing by 3.6% and 3.2%, respectively. In addition, the algorithm reduces the number of parameters by 26.6% and GFLOPs by 14.8%, realizing a lightweight design. These results confirm UES-YOLO’s effectiveness for underwater image target detection.</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

UES-YOLO: Underwater biological detection based on YOLO with an efficient sampling block

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Chaohao Shen,
  • Degang Yang,
  • Fei Liu,
  • Xin Zhang,
  • Liming Zhang

摘要

Underwater object recognition is crucial for achieving automated marine fishing and ensuring the sustainable development of marine resources. However, underwater environments’ complexity, such as low visibility and small, densely distributed targets, often causes detection issues like false positives, missed detections, and low accuracy. To tackle these challenges, this study introduces UES-YOLO, an enhanced lightweight underwater target detection algorithm based on YOLOv8n. We developed the Shallow and Maximum Attention Deep Robust Sampling Module (SMAD) to improve feature extraction. SMAD includes Shallow Robust Feature Downsampling (SRFD) for initial image processing and Maximum Attention Deep Robust Feature Downsampling (MADRFD) for deeper feature extraction. A Lightweight Asymmetric Detection Head (LADH) is also added to reduce parameters and computational complexity. The Minimum Point Distance Intersection over Union (MPDIoU) is used to enhance UES-YOLO’s generalization. Tests show UES-YOLO achieves mAP50 scores of 71% and 70.4% on the UDD and Aquarium datasets, improving by 8% and 1.8% over the original model. The mAP50-95 scores are 29.9% and 38.2%, increasing by 3.6% and 3.2%, respectively. In addition, the algorithm reduces the number of parameters by 26.6% and GFLOPs by 14.8%, realizing a lightweight design. These results confirm UES-YOLO’s effectiveness for underwater image target detection.