Scholarlink

<p>During the CNC machining, the blades exhibit various surface defects, including diverse morphologies and dimensions. Deep learning-based intelligent detection algorithms for the blade production line aim to improve computational efficiency and accuracy while minimizing model dimensions. This study proposes an enhanced blade detection method predicated upon a real-time detection transformer (RT-DETR) to detect blade surface defects precisely and efficiently in the blade production line. A dataset of blade surface defects in the blade machining process is first constructed, focusing on four surface defect types: gash, scratch, bruise, and pockmark. Secondly, the backbone network segment is substituted with an improved and more lightweight ResNet18 to optimize defect detection efficiency. The original feature fusion approach in RT-DETR is replaced by a Hierarchical Scale-based Feature Pyramid Network (HS-FPN) to enhance the model’s capability of detecting blade surface defects across various scales. The Inner-GIoU loss function is employed in RT-DETR to expedite model convergence and improve the accuracy of detecting minor surface defects. The results illustrate that the approach developed in this study raises the detection accuracy (mAP@0.5) by 3.5% and reduces the computational time required for detecting a single blade by 1.16&#xa0;s compared to the traditional RT-DETR. This algorithm exhibits a relatively faster detection speed and higher accuracy in the automated real-time detection of blade surface defects.</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

Real-time detection of blade surface defects based on the improved RT-DETR

  • 期刊论文

  • 发表日期：

    2025年2月8日

查看全文
  • Dongbo Wu,
  • Renkang Wu,
  • Hui Wang,
  • Zhijiang Cheng,
  • Suet To

摘要

During the CNC machining, the blades exhibit various surface defects, including diverse morphologies and dimensions. Deep learning-based intelligent detection algorithms for the blade production line aim to improve computational efficiency and accuracy while minimizing model dimensions. This study proposes an enhanced blade detection method predicated upon a real-time detection transformer (RT-DETR) to detect blade surface defects precisely and efficiently in the blade production line. A dataset of blade surface defects in the blade machining process is first constructed, focusing on four surface defect types: gash, scratch, bruise, and pockmark. Secondly, the backbone network segment is substituted with an improved and more lightweight ResNet18 to optimize defect detection efficiency. The original feature fusion approach in RT-DETR is replaced by a Hierarchical Scale-based Feature Pyramid Network (HS-FPN) to enhance the model’s capability of detecting blade surface defects across various scales. The Inner-GIoU loss function is employed in RT-DETR to expedite model convergence and improve the accuracy of detecting minor surface defects. The results illustrate that the approach developed in this study raises the detection accuracy (mAP@0.5) by 3.5% and reduces the computational time required for detecting a single blade by 1.16 s compared to the traditional RT-DETR. This algorithm exhibits a relatively faster detection speed and higher accuracy in the automated real-time detection of blade surface defects.