Scholarlink

<p>To address the limitations in weld image defect detection caused by complex backgrounds, blurred defect boundaries, and diverse defect morphologies, this paper proposes the RBPSD-YOLOv8 algorithm tailored for weld surface defect detection, which enhances detection performance through multi-faceted optimizations including replacing conventional Conv layers in the backbone and neck with RepConv modules to leverage multi-parameter branching structures for multi-scale feature extraction and reduce false detections for spatter and porosity defects while maintaining inference efficiency via reparameterization, introducing the Bottleneck Transformer (BoTNet Transformer) module to enhance the model’s capability to capture features of defects with varying shapes and sizes and address challenges from significant variability in weld-defect characteristics, adopting the Powerful-Iou version 2(PIoU v2) loss function in place of the original CIoU to accelerate model convergence and improve target localization accuracy, adding a dedicated small-target detection head optimized for micro-scale anomalies, such as porosity, cracks, and spatters, and replacing YOLOv8’s detection head with the lightweight DWDetect module to further boost detection efficiency. Experimental results demonstrate that the enhanced RBPSD-YOLOv8 model achieves improvements of 4.9% in precision (P), 3.9% in recall (R), and 5.0% in mean average precision (mAP), while reducing model parameters by 0.36M, computational complexity (GFLOPs) by 0.9 GFLOPs, and single-image processing time to 2.3ms, with comparative experiments on the NEU-DET and WELD-DEFECT.v1i.yolov8 datasets validating that the algorithm not only significantly enhances weld-defect detection performance but also exhibits strong generalization capabilities.</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

Rbpsd-yolov8: weld surface defect detection based on improved YOLOv8

  • 期刊论文

  • 发表日期：

    2025年7月22日

查看全文
  • Liuyi Ling,
  • Guo Wei,
  • Yuwen Liu,
  • Shuai Xu,
  • Liyu Wei,
  • Bolun Hong

摘要

To address the limitations in weld image defect detection caused by complex backgrounds, blurred defect boundaries, and diverse defect morphologies, this paper proposes the RBPSD-YOLOv8 algorithm tailored for weld surface defect detection, which enhances detection performance through multi-faceted optimizations including replacing conventional Conv layers in the backbone and neck with RepConv modules to leverage multi-parameter branching structures for multi-scale feature extraction and reduce false detections for spatter and porosity defects while maintaining inference efficiency via reparameterization, introducing the Bottleneck Transformer (BoTNet Transformer) module to enhance the model’s capability to capture features of defects with varying shapes and sizes and address challenges from significant variability in weld-defect characteristics, adopting the Powerful-Iou version 2(PIoU v2) loss function in place of the original CIoU to accelerate model convergence and improve target localization accuracy, adding a dedicated small-target detection head optimized for micro-scale anomalies, such as porosity, cracks, and spatters, and replacing YOLOv8’s detection head with the lightweight DWDetect module to further boost detection efficiency. Experimental results demonstrate that the enhanced RBPSD-YOLOv8 model achieves improvements of 4.9% in precision (P), 3.9% in recall (R), and 5.0% in mean average precision (mAP), while reducing model parameters by 0.36M, computational complexity (GFLOPs) by 0.9 GFLOPs, and single-image processing time to 2.3ms, with comparative experiments on the NEU-DET and WELD-DEFECT.v1i.yolov8 datasets validating that the algorithm not only significantly enhances weld-defect detection performance but also exhibits strong generalization capabilities.