基于YOLOv5s网络的光学玻璃曲面镜片缺陷检测方法
Optical glass curved lens based on YOLOv5s network Defect detection methods
刘小磊 1刘丰慧 1崔宸嘉2
作者信息
- 1. 河南理工大学物理与电子信息学院,河南焦作 454003
- 2. 郑州大学国际学院,郑州 450002
- 折叠
摘要
为了提升曲面光学玻璃镜片表面缺陷检测精度,对其表面缺陷自动化检测技术进行研究.通过分析光学玻璃镜片表面不同缺陷成像原理,设计一套两种打光方式相结合的缺陷采集装置,在采集到较高对比度缺陷图像的同时弥补光学镜片缺陷检测在脱膜缺陷上的不足;对采集到的缺陷图片进行预处理及数据增强操作,为光学玻璃镜片自动化缺陷检测提供高质量图片;将深度学习的方法运用在光学镜片缺陷检测上,通过对比不同网络模型在光学镜片缺陷数据集上的表现,选择效果最优的YOLOv5s完成对镜片缺陷的检测,其召回率和均值平均精度分别为92％和95％,检测一张缺陷镜片的时间为10 ms.
Abstract
In order to improve the accuracy of surface defect detection for curved optical glass lenses,automated surface defect detection technology is studied.By analyzing the imaging principles of different defects on the surface of optical glass lenses,a defect collection device combining two lighting methods is designed to capture high contrast de-fect images while compensating for the shortcomings of optical lens defect detection in detachment defects;Preprocess and enhance the collected defect images to provide high-quality images for automated defect detection of optical glass lens;Applying deep learning methods to optical lens defect detection,by comparing the performance of different net-work models on the optical lens defect dataset,Select the YOLOv5s with the best performance to complete the detection of lens defects,with a recall rate and average accuracy of 92％and 95％,respectively.The time to detect a defective lens is 10 ms.
关键词
缺陷检测/光学镜片/打光方式/深度学习Key words
defect detection/optical lenses/lighting method/deep learning引用本文复制引用
基金项目
国家自然科学基金面上项目(12074102)
出版年
2024
NETL
NSTL
万方数据