Scholarlink

<p>The accurate and efficient detection of bearing surface defects is critical for ensuring product quality and operational safety in industrial production environments. However, existing detection algorithms often suffer from insufficient lightweight design and limited detection accuracy, making them unsuitable for deployment on edge computing devices. To address these challenges, we propose ACS-YOLO, a novel lightweight object detection algorithm tailored for bearing surface defect detection. ACS-YOLO integrates four key modules: (1) the ADown downsampling module in both the Backbone and Neck, which reduces computational complexity and enhances defect feature extraction; (2) the C2f-RVB-EMA module, combining re-parameterized convolution and attention mechanisms to improve feature discrimination while maintaining model compactness; (3) a redesigned Neck incorporating the CSPHet architecture, which further reduces parameters and model size; and (4) a novel loss function FMIoU, which fuses Inner, Focaler-IoU and MPDIoU to mitigate the effects of class imbalance. Experimental results on multiple datasets show that ACS-YOLO achieves a mAP50 of 96.6%, while reducing FLOPs by 10.8%, parameters by 31.5%, and model size by 41.4% compared to the baseline.</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

ACS-YOLO: a lightweight bearing surface defect detection algorithm

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月1日

查看全文
  • Lei Ding,
  • Hong Xu,
  • Peng Du,
  • Yunhao Cui

摘要

The accurate and efficient detection of bearing surface defects is critical for ensuring product quality and operational safety in industrial production environments. However, existing detection algorithms often suffer from insufficient lightweight design and limited detection accuracy, making them unsuitable for deployment on edge computing devices. To address these challenges, we propose ACS-YOLO, a novel lightweight object detection algorithm tailored for bearing surface defect detection. ACS-YOLO integrates four key modules: (1) the ADown downsampling module in both the Backbone and Neck, which reduces computational complexity and enhances defect feature extraction; (2) the C2f-RVB-EMA module, combining re-parameterized convolution and attention mechanisms to improve feature discrimination while maintaining model compactness; (3) a redesigned Neck incorporating the CSPHet architecture, which further reduces parameters and model size; and (4) a novel loss function FMIoU, which fuses Inner, Focaler-IoU and MPDIoU to mitigate the effects of class imbalance. Experimental results on multiple datasets show that ACS-YOLO achieves a mAP50 of 96.6%, while reducing FLOPs by 10.8%, parameters by 31.5%, and model size by 41.4% compared to the baseline.