Scholarlink

<p>Resistance spot welding is a critical process in industrial production, where quality is paramount for ensuring product safety. Currently, the inspection of welded joint appearances relies on manual visual observation, which is characterized by slow speed and low efficiency, rendering it inadequate for modern automated welding production. This work introduces a lightweight inspection algorithm, YOLO-DMLite, designed to meet these requirements. Specifically, the YOLOv7-tiny detector is enhanced through the integration of the MobileNetv3 network, which is subsequently pruned using LAMP to achieve a lightweight design. Additionally, an ELAN-DBB structure is introduced in the detection head to mitigate accuracy loss associated with the lightweight network. The results demonstrate that the optimized YOLOv7 model exhibits a significant reduction in both parameter size and computational cost. Specifically, the model’s parameter size decreased from 23.04 to 3.77&#xa0;MB, and the computational complexity, measured in GFLOPS, was reduced from 13.1 to 2.8. Notably, the mAP@0.5&#xa0;achieved an impressive value of 96.4%.</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

Identification of appearance defects in resistance spot welding based on YOLOV7

  • 期刊论文

  • 发表日期：

    2025年8月26日

查看全文
  • Yuxi Zhu,
  • Yi Luo,
  • Yi Hu,
  • Xiaojun Deng,
  • Luohao Zhang

摘要

Resistance spot welding is a critical process in industrial production, where quality is paramount for ensuring product safety. Currently, the inspection of welded joint appearances relies on manual visual observation, which is characterized by slow speed and low efficiency, rendering it inadequate for modern automated welding production. This work introduces a lightweight inspection algorithm, YOLO-DMLite, designed to meet these requirements. Specifically, the YOLOv7-tiny detector is enhanced through the integration of the MobileNetv3 network, which is subsequently pruned using LAMP to achieve a lightweight design. Additionally, an ELAN-DBB structure is introduced in the detection head to mitigate accuracy loss associated with the lightweight network. The results demonstrate that the optimized YOLOv7 model exhibits a significant reduction in both parameter size and computational cost. Specifically, the model’s parameter size decreased from 23.04 to 3.77 MB, and the computational complexity, measured in GFLOPS, was reduced from 13.1 to 2.8. Notably, the mAP@0.5 achieved an impressive value of 96.4%.