基于改进YOLOv8n的PCB缺陷检测算法  

PCB defect detection algorithm based on improved YOLOv8n

作　　者：姜源 付波[1] 权轶[1] 李昊 Jiang Yuan;Fu Bo;Quan Yi;Li Hao(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China)

机构地区：[1]湖北工业大学电气与电子工程学院,武汉430068

出　　处：《国外电子测量技术》2024年第6期22-32,共11页Foreign Electronic Measurement Technology

基　　金：湖北省重点研发计划(2021BAA193)项目资助。

摘　　要：针对现有的印刷电路板(PCB)缺陷检测方法计算量大、小目标缺陷易漏检、检测速度较慢等问题,提出YOLOv8n-4SCDP缺陷检测算法。首先,在YOLOv8n颈部网络增加上采样,融合Backbone中浅层语义信息,同时增加微小目标检测层降低PCB小目标缺陷漏检率;其次,在Backbone中融入坐标注意力(CA)机制,强化特征语义和位置信息,提高了模型特征融合能力;另外,设计密集连接机构,提高模型的缺陷特征利用率,采用PConv对模型进行压缩,既保证了模型的准确性,又大大减小了模型的尺寸;最后,针对难易样本不平衡的问题,采用线性区间映射法重新定义回归损失函数(Focaler-SIoU),提高模型收敛速度和回归精度。实验结果表明,YOLOv8n-4SCDP算法的整体缺陷的平均精度均值(mAP)达到95.8%,检测帧率达到了65fps。有效改善YOLOv8n对于PCB小目标缺陷漏检率高、检测精度低等问题。In response to the issues of large computational load,easy omission of small target defects,and slow detection speed in existing PCB defect detection methods,this paper proposes the YOLOv8n-4SCDP defect detection algorithm.Firstly,upsampling is added to the neck network of YOLOv8n,integrating shallow semantic information in the Backbone,and a small target detection head is added to reduce the omission rate of small target defects in PCBs.Secondly,the CA attention mechanism is integrated into the Backbone to enhance the semantic and positional information of features,thereby improving the feature fusion capability of the model.Thirdly,a dense connection mechanism was designed to enhance the utilization of defect features in the model.Additionally,PConv was employed to compress the model,ensuring both accuracy and significantly reducing the model’s size.Finally,to address the issue of imbalanced difficult and easy samples,we employ a linear interval mapping method to redefine the Focaler-SIoU regression loss function.This approach enhances both model convergence speed and regression accuracy.The experimental results indicate that the YOLOv8n-4SCDP algorithm achieves an accuracy of 95.8%and a frame rate of 65 fps.This effectively addresses YOLOv8n’s issues related to high defect omission rates and low detection accuracy for small PCB targets.

关 键 词：YOLOv8n PCB缺陷 小目标缺陷检测 密集连接 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]