基于YOLOv4改进的PCB板缺陷检测算法  

Improved PCB Board Defect Detection Algorithm Based on YOLOv4

作　　者：李致金[1] 江凯强 高伟 刘忠洋 LI Zhijin;JIANG Kaiqiang;GAO Wei;LIU Zhongyang(School of Artificial Intelligence(School of Future Technology),Nanjing University of Information Science&Technology,Nanjing 210044)

机构地区：[1]南京信息工程大学人工智能学院(未来技术学院),南京210044

出　　处：《计算机与数字工程》2025年第2期320-326,共7页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61971167)资助。

摘　　要：针对当前工业中PCB板缺陷检测方法存在精度低、推理速度慢与模型体积大等问题,现提出一种基于YO-LOv4改进的PCB板缺陷检测方法。首先将YOLOv4主干网络替换为GhostNet网络,大大降低主干特征提取网络的参数量,缩小模型体积,其次在主干网络中加入GCT注意力机制,在不增加计算复杂度的情况下强化特征提取能力,提高精度,最后使用蓝图卷积,降低算法计算复杂度的同时提高检测精度,实现轻量化。使用北京大学智能机人开放实验室公开的PCB瑕疵数据集进行实验,实验结果表明,所提改进算法轻量、高效,对比原算法,在mAP精度与检测速度上均有提升,模型大小降低,能够解决当前存在的问题。Aiming at the problems of low precision,slow inference speed and large model size in the current PCB defect detec⁃tion method in the industry,an improved PCB defect detection method based on YOLOv4 is proposed.Firstly,the YOLOv4 back⁃bone network is replaced with a GhostNet network,which greatly reduces the number of parameters of the backbone feature extrac⁃tion network and reduces the size of the model.Secondly,the GCT attention mechanism is added to the backbone network to en⁃hance feature extraction capabilities without increasing computational complexity,improve the accuracy,and finally use the blue⁃print convolution to reduce the computational complexity of the algorithm and improve the detection accuracy to achieve lightweight.Using the PCB defect data set published by the intelligent robot open laboratory of Peking university to conduct experiments,the ex⁃perimental results show that the proposed improved algorithm is lightweight and efficient.Compared with the original algorithm,the mAP accuracy and detection speed are improved,and the model size is reduced,it can solve the current problems.

关 键 词：缺陷检测 YOLOv4 GhostNet GCT注意力 蓝图卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]