Scholarlink

<p>Hydraulic tunnels at hydropower stations, essential for water release and energy reduction, frequently degrade due to multiple stressors. Despite the widespread use of forward-looking sonar for inspections, its accuracy is often insufficient to address the complex surface defects and demanding underwater circumstances in these tunnels. To address these challenges, a dynamic snake convolution enhanced YOLOv8s model, named DSW-YOLOv8s, is proposed to detect hydraulic tunnel defect. Firstly, acknowledging the slender and subtle nature of hydraulic tunnel defects, dynamic snake shaped convolution (DSConv) is introduced to the Neck section of YOLOv8s network. This enhancement enables it to precisely capture the characteristics of tubular structures by adaptively focusing on slender and curved features. Secondly, to overcome the difficulties posed by low contrast and small lesion area in sonar images, a dedicated small object detection head is integrated into the YOLOv8s network, enhancing its ability to detect small objects. Finally, the WIoUv3 is employed for bounding box regression loss, effectively amplifying gradient gains for samples of varying quality and improving positioning accuracy. Experiments conducted on a self-built dataset demonstrate that DSW-YOLOv8s outperforms both the original YOLOv8s network and other mainstream deep neural networks in the detection of hydraulic tunnel defects, achieving an mAP@0.5 of 88.7% and an mAP@0.5:0.95 of 49.2%. The source code and dataset of DSW-YOLOv8s can be viewed at <a href="https://github.com/WJX192443/DSW-YOLOv8s.git">https://github.com/WJX192443/DSW-YOLOv8s.git</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

Dynamic snake convolution enhanced YOLOv8s for hydraulic tunnel defect detection

  • 期刊论文

  • 发表日期：

    2025年5月29日

查看全文
  • Jingxian Wang,
  • Longgang Zhang,
  • Wang Peng,
  • Shuangshuang Ma,
  • Chenhao Zhu

摘要

Hydraulic tunnels at hydropower stations, essential for water release and energy reduction, frequently degrade due to multiple stressors. Despite the widespread use of forward-looking sonar for inspections, its accuracy is often insufficient to address the complex surface defects and demanding underwater circumstances in these tunnels. To address these challenges, a dynamic snake convolution enhanced YOLOv8s model, named DSW-YOLOv8s, is proposed to detect hydraulic tunnel defect. Firstly, acknowledging the slender and subtle nature of hydraulic tunnel defects, dynamic snake shaped convolution (DSConv) is introduced to the Neck section of YOLOv8s network. This enhancement enables it to precisely capture the characteristics of tubular structures by adaptively focusing on slender and curved features. Secondly, to overcome the difficulties posed by low contrast and small lesion area in sonar images, a dedicated small object detection head is integrated into the YOLOv8s network, enhancing its ability to detect small objects. Finally, the WIoUv3 is employed for bounding box regression loss, effectively amplifying gradient gains for samples of varying quality and improving positioning accuracy. Experiments conducted on a self-built dataset demonstrate that DSW-YOLOv8s outperforms both the original YOLOv8s network and other mainstream deep neural networks in the detection of hydraulic tunnel defects, achieving an mAP@0.5 of 88.7% and an mAP@0.5:0.95 of 49.2%. The source code and dataset of DSW-YOLOv8s can be viewed at https://github.com/WJX192443/DSW-YOLOv8s.git.