YOLOv5改进的印制电路板缺陷检测  

Improved Defect Detection for Printed Circuit Board Based on YOLOv5

作　　者：李华洲[1] 张青[1] 杨锦鸿 LI Hua-zhou;ZHANG Qing;YANG Jin-hong(Nanning College For Vocational Technology,Nanning,Guangxi,China 530000)

机构地区：[1]南宁职业技术学院,广西南宁530000

出　　处：《湖南邮电职业技术学院学报》2023年第1期32-36,共5页Journal of Hunan Post and Telecommunication College

基　　金：2018年广西职业教育第一批专业发展研究基地“广西职业教育新一代信息技术专业群发展研究基地”(项目编号:桂教职成[2018]37号03);2022年度广西高校中青年教师科研基础能力提升项目“基于深度学习的PCB裸板缺陷检测系统研究”(项目编号:2022KY1018)。

摘　　要：针对印制电路板缺陷检测因检测目标相对较小而产生的检测精确率不高的问题,以及基于深度学习算法模型目标检测对算力的高负荷,提出一种基于YOLOv5的改进深度学习网络结构,通过增加特征提取层和特征融合层提升网络对小目标检测的能力。同时对训练得到的模型采用模型裁剪、模型量化将模型优化,并通过模型校准在尽可能保留缺陷检测精确度的同时减小网络规模。实验分析改进后的模型检测精确度可达到96.2%,而压缩优化后的模型精确度也可以达到94.4%,而模型大小可以减小73.7%,其性能指标满足实际应用中印制电路板缺陷检测的要求。To solve the problem of low detecting rate of accuracy when detecting the defects on printed circuit board due to the small size objects of defect,and to tackle the high burden of computation of model object detection based on the deep learning algorithm,this paper provides an improved network structure of convolution neural network,which improves the detecting ability of network by adding feature attracting layer and feature concatenating layer to the network.Meanwhile,this paper utilizes model prune,model quantization,and model calibration to reduce the scale of network and keep the detecting accuracy as much as possible.Through experiment it can be seen that the improved model can reach 96.2% of accuracy rate of detection,and the compressed model can still maintain the high accuracy rate which is 94.4%,but the size of model declines 73.7%,which is satisfied for the practical application of the requirement when detecting defects on printed circuit board.

关 键 词：深度学习 印制电路板 目标检测 模型裁剪 模型量化 模型校准 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]