基于深度学习的PCB焊锡缺陷检测  

Research on PCB Solder Defect Defection Based on Deep Learning

作　　者：卢小康 欧阳华兵[1] 陈田[1] 刘军[2] LU Xiaokang;OUYANG Huabing;CHEN Tian;LIU Jun(School of Mechanical Engineering,Shanghai Dianji University,Shanghai 201306,China;Kaiserslautern Kolleg fur Intelligente Produktion,Shanghai Dianji University,Shanghai 201306,China)

机构地区：[1]上海电机学院机械学院,上海201306 [2]上海电机学院凯劳智能制造学院,上海201306

出　　处：《无线电工程》2024年第2期276-283,共8页Radio Engineering

基　　金：上海市地方院校能力建设计划项目(22010501000);上海多向模锻工程技术研究中心资助项目(20DZ2253200);上海市临港新片区智能制造产业学院资助项目(B1-0299-21-023)。

摘　　要：针对现有印刷电路板(PCB)缺陷检测算法具有体积大、精度差和漏检率高等问题,提出一种改进的YOLOv5s算法,利用迁移学习进行优化。在主干特征提取网络加入卷积块注意力模块(Convolutional Block Attention Module,CBAM)并引出一个新的蕴含更多丰富细节信息的浅层特征来提高模型信息感知能力,增强网络对小目标的检测。使用C3-CBAM替换加强特征提取网络的C3结构,为了在保障模型检测精度的同时减少模型参数量,使用深度可分离卷积替换下采样。去掉用于大目标的检测头(Yolo Head),避免数据不均衡带来的先验框分配问题。在自制PCB焊锡缺陷数据集上实验表明,改进后算法较原YOLOv5s算法的模型参数量减少24.8%,平均精度均值(mean Average Precision,mAP)达到99.46%,较原YOLOv5s提升了5.45%,证明了改进措施的有效性。For existing PCB defect detection algorithms with large size,poor accuracy and high missing rate problems,an improved YOLOv5s algorithm is proposed and optimized by using migration learning.Firstly,a Convolutional Block Attention Module(CBAM)is added to the backbone feature extraction network,and a new shallow feature with more detailed information is introduced to improve the information perception of the model and enhance the detection of small targets.Secondly,the C3 structure of the enhanced feature extraction network is replaced by C3-CBAM.To reduce the number of model parameters while maintaining model detection accuracy,depth-separable convolution is used to replace down-sampling.Lastly,the detection head(Yolo Head)used for large targets is removed to avoid the problem of a priori frame assignment due to data imbalance.Experiments on a home-made PCB solder defect dataset show that the improved algorithm reduces the number of model parameters by 24.8%compared to the original YOLOv5s algorithm and achieves a mean Average Precision(mAP)of 99.46%,a 5.45%improvement over the original YOLOv5s,demonstrating the effectiveness of the improvement measures.

关 键 词：印刷电路板 YOLOv5s 缺陷检测 小目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]