Scholarlink

<p>To address the challenges in ship welding defect detection, such as small defect targets, complex backgrounds that hinder feature extraction, and uneven attention distribution, this paper proposes a precise and lightweight YOLOv8-RSN-LPD model based on the YOLOv8 framework, aiming to improve the accuracy and robustness of ship welding defect recognition while maintaining a lightweight design. First, in terms of feature extraction, the RepGhost backbone is introduced. By employing multi-branch convolutions and structural re-parameterization, fine-grained feature representation is enhanced, effectively improving the recognition capability for small defect targets. Second, a novel MultiResSimAM attention mechanism is proposed and integrated into both the backbone and neck networks. This mechanism leverages multi-scale processing and residual structures to aggregate contextual information, significantly improving feature focus and localization accuracy for small targets. Furthermore, to address the insufficient sensitivity to tiny targets, the Normalized Wasserstein Distance (NWD) loss function is adopted, effectively enhancing the stability of bounding box regression and the robustness of detection. Finally, a lightweight LPD module is designed in the detection head, further optimizing the model architecture through partial convolutions. Compared with the original YOLOv8n model, the proposed YOLOv8-RSN-LPD achieves a Precision of 0.941 and mAP@50 of 0.96, increasing by 4.8% and 1.2%, respectively, while reducing the number of parameters, GFLOPs, and model size by 25.58%, 30.49%, and 19.0%, fully demonstrating the effectiveness and practicality of the proposed method.</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

A lightweight YOLOv8 detection algorithm for X-ray ship welding defects in complex backgrounds

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Caiping Liang,
  • Chenjun Xu,
  • Wenxu Niu,
  • Cui Li

摘要

To address the challenges in ship welding defect detection, such as small defect targets, complex backgrounds that hinder feature extraction, and uneven attention distribution, this paper proposes a precise and lightweight YOLOv8-RSN-LPD model based on the YOLOv8 framework, aiming to improve the accuracy and robustness of ship welding defect recognition while maintaining a lightweight design. First, in terms of feature extraction, the RepGhost backbone is introduced. By employing multi-branch convolutions and structural re-parameterization, fine-grained feature representation is enhanced, effectively improving the recognition capability for small defect targets. Second, a novel MultiResSimAM attention mechanism is proposed and integrated into both the backbone and neck networks. This mechanism leverages multi-scale processing and residual structures to aggregate contextual information, significantly improving feature focus and localization accuracy for small targets. Furthermore, to address the insufficient sensitivity to tiny targets, the Normalized Wasserstein Distance (NWD) loss function is adopted, effectively enhancing the stability of bounding box regression and the robustness of detection. Finally, a lightweight LPD module is designed in the detection head, further optimizing the model architecture through partial convolutions. Compared with the original YOLOv8n model, the proposed YOLOv8-RSN-LPD achieves a Precision of 0.941 and mAP@50 of 0.96, increasing by 4.8% and 1.2%, respectively, while reducing the number of parameters, GFLOPs, and model size by 25.58%, 30.49%, and 19.0%, fully demonstrating the effectiveness and practicality of the proposed method.