基于VSD-YOLOv5s的轻量化注塑齿轮缺陷检测
Defect Detection of Lightweight Injection Molding Gear Based on VSD-YOLOv5s
申飞 1周敏 1黄周林 1李鑫炎 1张美洲1
作者信息
- 1. 武汉科技大学冶金装备及其控制教育部重点实验室,武汉 430081;武汉科技大学机械传动与制造工程湖北省重点实验室,武汉 430081;武汉科技大学精密制造研究院,武汉 430081
- 折叠
摘要
针对注塑齿轮加工过程中在线缺陷检测存在的识别速度较慢、识别精度较低等问题,提出了一种基于ShuffleNetV2 主干网络的改进YOLOv5s网络模型.在YOLOv5s网络模型的基础上将原来的C3 主干网络结构替换为轻量化ShuffleNetV2 结构,降低模型参数量和尺寸,提升识别速度.引入SE注意力机制,并通过DIOU-NMS方法去除冗余框,减少错误抑制,提高识别精度.实验结果表明:相比于原模型,改进的模型识别准确率提升了0.9%,平均识别精度提升了1.7%,识别速度提升了5 fps,满足注塑齿轮表面缺陷在线识别的速度以及精度要求.
Abstract
An improved YOLOv5s network model based on ShuffleNetV2 backbone network was proposed after analyzing the problems of slow identification speed and low accuracy of online defect detection in in-jection molding gear processing.On the basis of YOLOv5s network model,the original C3 backbone net-work structure is replaced by lightweight ShuffleNetV2 structure,reducing the number and size of model parameters and improving the recognition speed.The SE attention mechanism is introduced,and the DIOU-NMS method is used to remove redundant boxes,reduce error suppression,and improve recognition accura-cy.The experimental results show that compared with the original model,the recognition accuracy of the improved model is increased by 0.9%,the average recognition accuracy is increased by 1.7%,and the rec-ognition speed is increased by 5 fps,which can meet the speed and accuracy requirements of online recogni-tion of injection gear surface defects.
关键词
ShuffleNetV2/YOLOv5s/注塑齿轮/缺陷检测Key words
ShuffleNetV2/YOLOv5s/injection molding gear/defect detection引用本文复制引用
基金项目
国家自然科学基金项目(51975431)
武汉科技大学研究生创新创业项目(JCX2021048)
出版年
2024
NETL
NSTL
维普
万方数据