Scholarlink

<p>The detection of surface defects in industrial products is vital for ensuring product quality. Ensuring real-time performance, while improving the detection accuracy of low-pixel-resolution small defects against background interference poses a significant challenge. To address this, the EA-YOLO model, based on Yolov8, is proposed. It includes three main improvements: replacing C2f (Faster Implementation of CSP Bottleneck with 2 convolutions) with a specially designed C2FN (Faster Implementation of CSP FastNet Block with 2 convolutions Network) in the backbone module to reduce parameters and GFLOPs, while enhancing speed; introducing the Environmental Awareness Dynamic Network (EADN) to prevent the loss of defect information in extreme positions; and using the improved Dynamic Adaptive Fusion Detector (DAF-Detect) for predictions. Case studies with the NEU-DET and PCB-DET datasets show that EA-YOLO achieves a mAP of 81.1 and 97.8%, respectively, improving by 4.3 and 4.1% compared to the baseline model, with reduced parameters, GFLOPs, and increased FPS, demonstrating good robustness and generalization ability.</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 small defect detection technique for industrial product surfaces based on the EA-YOLO model

  • 期刊论文

  • 发表日期：

    2025年1月21日

查看全文
  • Biao Li,
  • Bing Wang,
  • Xiong Hu,
  • Jianhui Zhai,
  • Changping Ji

摘要

The detection of surface defects in industrial products is vital for ensuring product quality. Ensuring real-time performance, while improving the detection accuracy of low-pixel-resolution small defects against background interference poses a significant challenge. To address this, the EA-YOLO model, based on Yolov8, is proposed. It includes three main improvements: replacing C2f (Faster Implementation of CSP Bottleneck with 2 convolutions) with a specially designed C2FN (Faster Implementation of CSP FastNet Block with 2 convolutions Network) in the backbone module to reduce parameters and GFLOPs, while enhancing speed; introducing the Environmental Awareness Dynamic Network (EADN) to prevent the loss of defect information in extreme positions; and using the improved Dynamic Adaptive Fusion Detector (DAF-Detect) for predictions. Case studies with the NEU-DET and PCB-DET datasets show that EA-YOLO achieves a mAP of 81.1 and 97.8%, respectively, improving by 4.3 and 4.1% compared to the baseline model, with reduced parameters, GFLOPs, and increased FPS, demonstrating good robustness and generalization ability.