首页|基于轻量化改进型YOLOv5s的PCB裸板缺陷检测

基于轻量化改进型YOLOv5s的PCB裸板缺陷检测

扫码查看
提高印刷电路板(printed circuit board,PCB)裸板缺陷检测能力对智能制造生产具有重大意义.本文提出一种轻量化改进型YOLOv5s的PCB裸板缺陷检测算法.该方法采用轻量型的Mobilenetv3的特征提取模块取代YOLOv5s的主干结构,大大降低了模型的体积大小,且减小了其浮点运算次数,充分节省了设备运行硬件资源.针对PCB裸板缺陷特征体积形态复杂、分布不均匀的特点,增加了坐标注意力模块和160×160大小的特征检测头,进一步提高了 PCB裸板缺陷检测精确度.通过实验验证得出:改进后的网络模型对比于原YOLOv5s模型,其模型体积大小为原网络模型的30.5%,而浮点运算次数仅为原网络模型的29.8%,并且其对PCB裸板缺陷检测的精确率达98.15%,召回率达98.95%,mAP0.5及mAP0.5∶0.95也都达到了98.35%和65.79%.改进后的网络模型有助于为嵌入式等硬件资源有限设备中PCB裸板缺陷检测技术的开发提供技术支持.
PCB Bare Board Defect Detection Based on Lightweight Improved YOLOv5s
Improving the detection ability of bare board defects in printed circuit board is of great significance for intelligent manufacturing production.This article proposes a lightweight improved YOLOv5s PCB bare board defect detection algorithm.This method uses a lightweight Mobilenetv3 feature extraction module to replace the backbone structure of YOLOv5s,greatly reduced the size of the model and the number of floating-point operations,fully saved hardware resources.In response to the complex distribution of PCB bare board defect features,a coordinate attention module and additional 160 X 160 size feature detection head has been introduced to improve the accuracy of PCB bare board defect detection.Through experiment,it was found that compared with the original YOLOv5s model,the size of improved network model is only 30.5%of the original network model,and floating-point operation is only 29.8%of the original network model.Moreover,its accuracy in detecting PCB bare board defects reaches 98.15%,and its recall rate reaches 98.95%.Both mAP0.5 and mAP0.5∶0.95have reached 98.35%and 65.79%.The improved network model helps to provide technical support for the development of PCB bare board defect detection technology in embedded and other hardware resource limited devices.

lightweightYOLOv5scoordinate attentiondefect detection

计甜甜、黄晓姣、汪明珠、黄济

展开 >

皖西学院实验实训教学管理部,安徽六安 237012

轻量级 YOLOv5s 坐标注意力机制 缺陷检测

皖西学院校级自然青年项目皖西学院校级自然青年项目安徽省教育厅高等学校省级质量工程项目

WXZR201914WXZR2020092022jcjs129

2024

皖西学院学报
皖西学院

皖西学院学报

影响因子:0.299
ISSN:1009-9735
年,卷(期):2024.40(2)
  • 11