摘要
针对现存印刷电路板(PCB)缺陷检测方法计算复杂、模型参数量大,不能满足轻量化部署要求的问题,提出基于轻量化YOLOv5的PCB表面缺陷检测方法。该方法使用重参数化视觉几何组(RepVGG)模块组成特征提取网络,解决YOLOv5中因主干网络中参数过多,从而难以部署到移动端的问题;在颈部网络中,采用全局交互卷积(GSConv)替换了部分卷积,保持了模型的精度,使输出更接近标准卷积;此外,采用有效交并比(EIOU)损失函数,加快模型的收敛速度,并提高模型的检测精度。结果表明:相较于YOLOv5,该方法使得参数量减少28.7%,计算量降低13.3%,平均精度稳定在99%以上。该网络的提出为PCB的缺陷检测问题提供了一种更高效的方法。
A printed circuit board(PCB)surface defect detection method based on lightweight YOLOv5 is proposed to address the problems of complex calculation and large model parameters that do not meet the requirements of lightweight deployment in existing PCB defect detection methods.This method uses a re-parameterization visual geometry group(RepVGG)module to form a feature extraction network to solve the problem that YOLOv5 is difficult to deploy to the mobile terminal due to too many parameters in the backbone network.In the neck network,the global shuffle convolution(GSConv)is used to replace some convolutions to keep the accuracy of the model and make the output closer to the standard convolution.In addition,the efficient intersection over union(EIOU)loss function is applied to accelerate the convergence speed of the model and improve the detection accuracy of the model.The results show that compared with YOLOv5,this method reduces the number of parameters by 28.7%,reduces the computational complexity by 13.3%,and stabilizes the average accuracy above 99%.The proposed network provides a more efficient method for PCB defect detection.
作者
余丽娜
张栋(指导)
俞雪锋
何智
许泽林
YU Lina;ZHANG Dong;YU Xuefeng;HE Zhi;XU Zelin(School of Mechanical Engineering,Shanghai Dianji University,Shanghai 201306,China;R&D Department,Shanghai Yihang Auto Parts Co.,Ltd.,Shanghai 201316,China)
出处
《上海电机学院学报》
2023年第4期226-231,238,共7页
Journal of Shanghai Dianji University
作者简介
余丽娜(1998-),女,硕士生,主要研究方向为图像处理、目标检测,E-mail:1418534428@qq.com;张栋(1968-),男,副教授,博士,主要研究方向为机电液一体化技术,E-mail:zhangd@sdju.edu.cn。
