改进YOLOv7的PCB缺陷检测算法  

An Improved YOLOv7 Algorithm for PCB Defect Detection

作　　者：王玲 向北平[1] 张晓勇[1] WANG Ling;XIANG Beiping;ZHANG Xiaoyong(School of Manufacturing Science and Engineering,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China)

机构地区：[1]西南科技大学制造科学与工程学院,四川绵阳621010

出　　处：《机械科学与技术》2025年第1期9-18,共10页Mechanical Science and Technology for Aerospace Engineering

基　　金：四川省科技厅重点研发计划(23ZDYF0471)。

摘　　要：针对检测印刷电路板(Printed circuit board, PCB)缺陷任务中,通用物体检测算法难以区分目标缺陷与背景,从而导致检测精度低等问题,提出一种改进YOLOv7的PCB表面缺陷检测模型。首先,在主干提取网络用Conv2Former(Transformer-style convolutional network)模块替代ELAN模块,保留空间信息的同时加强全局信息关联性,有效减少参数量。其次,删除20×20的大目标检测层,增加160×160的小目标检测层,以此保留更多小目标信息。此外,在特征融合网络引入SimAM(Similarity-based attention mechanism)注意力机制,不引入额外参数的同时提升检测精确度。最后,将Focal损失函数与CIoU损失函数结合,优化损失函数中高质量与低质量样本的权重分配,提升检测效果。实验结果表明,改进后的模型平均检测精度达到95.3%,相较于原模型精度提高了3.6%,参数量为10.97 MB,仅为原模型参数量的三分之一,改进后的模型能够更准确地识别PCB缺陷,有效降低漏检和误检率。In the task of detecting defects on printed circuit boards(PCB),general object detection algorithms always struggle to distinguish target defects from the background,resulting in issues such as low detection accuracy.In order to solve these problems,an improved YOLOv7 model for PCB surface defect detection is proposed.Firstly,the ELAN module is replaced with the transformer-style convolutional network(Conv2Former)module in the backbone extraction network,which preserves spatial information,strengthens global correlations,and effectively reduces the number of parameters.Secondly,to retain more information on small targets,the 20×20 layer is removed,and a 160×160 layer is added.Additionally,the introduction of the similarity-based attention mechanism(SimAM)in the feature fusion network enhances detection accuracy without introducing additional parameters.Finally,the Focal-CIoU Loss function,a combination of Focal Loss and CIoU Loss,optimizes weight allocation for high-quality and low-quality samples,and the detection effectiveness is enhanced.The improved YOLOv7 PCB surface defect detection algorithm achieves a mean average precision(mAP)of 95.3%,a 3.6%boost over the original model,with just 10.97 MB parameters and only a third of the original model.This refined model identifies PCB defects more precisely,significantly reducing leakage and false detections.

关 键 词：PCB表面缺陷检测 YOLOv7 Conv2Former SimAM Focal-CIoU 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TN41[自动化与计算机技术—计算机科学与技术]