东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (7): 1002-1010.DOI: 10.12068/j.issn.1005-3026.2024.07.012

• 机械工程 • 上一篇    下一篇

基于改进YOLOv5的退役轴类零件表面损伤检测方法

刘伟嵬(), 邱佳鹤, 胡光大, 刘泽远   

  1. 大连理工大学 高性能精密制造全国重点实验室,辽宁 大连 116024
  • 收稿日期:2023-03-20 出版日期:2024-07-15 发布日期:2024-10-29
  • 通讯作者: 刘伟嵬
  • 基金资助:
    国家自然科学基金资助项目(52175455)

Surface Damage Detection Method for Retired Shaft Parts Based on Improved YOLOv5

Wei-wei LIU(), Jia-he QIU, Guang-da HU, Ze-yuan LIU   

  1. State Key Laboratory of High-Performance Precision Manufacturing,Dalian University of Technology,Dalian 116024,China.
  • Received:2023-03-20 Online:2024-07-15 Published:2024-10-29
  • Contact: Wei-wei LIU
  • About author:LIU Wei-weiE-mail:liuww@dlut.edu.cn

摘要:

针对传统检测方法在对退役轴类零件进行损伤检测时存在效率低、结果一致性差等问题,提出了一种基于改进YOLOv5的退役轴类零件表面损伤检测方法.首先,将注意力机制嵌入检测算法中,增强了损伤在图像中的特征表示;然后,采用重复加权双向特征融合方法改进了检测模型的网络结构,有效提升了网络特征提取能力;最后,使用Ghostconv卷积模块代替普通卷积,大幅度降低了模型参数量.实验结果显示,改进后的算法模型精度比原始YOLOv5提升了6.9%,达到88.4%,同时模型参数量减少了6.1%,保证了检测速度与YOLOv5持平.与YOLOv3,SSD,Faster-RCNN等主流检测方法相比,在保证较高检测速度的同时,检测精度也有着明显优势.

关键词: YOLOv5, 表面损伤检测, 注意力机制, 多路特征融合, Ghostconv

Abstract:

Aiming at the existing problems of low efficiency and poor consistency in the damage detection of retired shaft parts by traditional detection methods, an improved YOLOv5?based surface damage detection method for retired shaft parts is proposed. Firstly, the attention mechanism is embedded into the detection algorithm to enhance the feature representation of damage in the image. Then the network structure of the detection model is improved by using the repeated weighted bidirectional feature fusion method to effectively enhance the network feature extraction capability. Finally, Ghostconv convolution module is used instead of normal convolution, which drastically reduces the number of model parameters. The experimental results show that the accuracy of the modified algorithm model has improved by 6.9% compared to the original YOLOv5, reaching 88.4%, while the number of model parameters has reduced by 6.1%, ensuring the detection speed is on par with YOLOv5. Compared with such mainstream detection methods as YOLOv3, SSD and Faster-RCNN, its detection accuracy has a significant advantage while ensuring a higher detection speed.

Key words: YOLOv5, surface damage detection, attention mechanism, multiplex feature fusion, Ghostconv

中图分类号: