基于改进U-Net网络的PCB缺陷检测方法  

PCB Defect Detection Method Based on Improved U-Net Network

作　　者：彭勇[1] 刘慧民 李伟松 王石 PENG Yong;LIU Huimin;LI Weisong;WANG Shi(Electronic Information School,Dongguan Polytechnic,Dongguan,Guangdong 523808,China;Dongguan Somacis Graphic PCB Co.,Ltd.,Dongguan,Guangdong 523800,China;Dongguan Huaxin Intelligent Technology Co.,Ltd.,Dongguan,Guangdong 523808,China)

机构地区：[1]东莞职业技术学院电子信息学院,广东东莞523808 [2]东莞森玛仕格里菲电路有限公司,广东东莞523800 [3]东莞华信智能技术有限公司,广东东莞523808

出　　处：《计算技术与自动化》2025年第1期183-188,共6页Computing Technology and Automation

基　　金：广东省教育厅2023年度普通高校重点领域项目(2023ZDZX1087);东莞科技特派员项目(20221800500812);东莞职业技术学院国家双高计划电子信息工程技术专业群专项政校行企项目(ZXD202302)。

摘　　要：针对PCB表面小尺寸缺陷难以检测的问题,提出了一种改进的U-Net语义分割网络,实现PCB表面缺陷图像的精确检测。首先,将U-Net的四层网络层次修改为三层,可以减少整体的计算工作量、提升网络模型收敛速度、缩短训练时间;其次,在U-Net网络中融入CBAM(Convolutional Block Attention Module)模块来提升图像中缺陷目标的显著度;然后,在编码阶段使用混合空洞卷积替换原有卷积块,增大感受野,获取更多的上下文信息。结果表明,U-Net的改进模型能够在提升模型性能的同时减少计算复杂度,能够增加PCB缺陷检测效率。Aiming at the low detection efficiency of the small-size defects of PCB surface,which cannot meet the real time detection requirements industrial production,a defect recognition method based on improved U-Net for printed circuit board defect detection is proposed.Firstly,reduce the original four layers of U-Net to three layers to reduce network computation and shorten model training time.Secondly,the convolutional block attention module(CBAM)was integrated into the U-Net network to improve the significance of the defective targets in the image.Finally,in the encoding stage,the hybrid dilated convolution is used to replace the original convolution block to increase the receptive field and obtain more context information.The results show that the improved U-Net model network improves network performance while reducing the computational complexity of the U-Net network and can increase the efficiency of printed circuit board defect detection.

关 键 词：缺陷检测 U-Net 空洞卷积 注意力机制 语义分割网络 轻量型网络 深度学习 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]