基于改进YOLOv5网络的印刷电路板缺陷检测  

PCB Defect Detection Based on Improved YOLOv5 Network

作　　者：卢麒仰 徐超林 杨育 伍梓帆 赵洪琛 刘宏展[1] LU Qiyang;XU Chaolin;YANG Yu;WU Zifan;ZHAO Hongchen;LIU Hongzhan(School of Information Optoelectronic Technology,South China Normal University,Guangzhou Guangdong 510631,China)

机构地区：[1]华南师范大学信息光电子科技学院,广东广州510631

出　　处：《电子器件》2023年第6期1504-1508,共5页Chinese Journal of Electron Devices

基　　金：国家自然科学基金项目(62175070,61875057);华南师范大学大学生创新创业项目(202132012)。

摘　　要：针对印刷电路板(PCB)现存检测方法效率低和对小缺陷目标检测准确率低等缺点,提出针对YOLOv5网络进行改进,通过增加小目标检测层获取更多小缺陷特征,之后增加FPN算法融合深浅层的特征信息,提高深层的特征综合度。同时通过图像分割操作放大缺陷占比,提高精确度。结果表明,相比于优化前,设计的检测系统对PCB图片的检测准确性提高了1.89%,检测的缺陷平均检测精度均值提高了1.82%,并且减少了非必要的检测。这为完善PCB的高效检测提供了一定参考。In view of the low efficiency of existing detection methods for printed circuit board(PCB)and low accuracy of small defect target detection,the YOLOv5 network is proposed to be improved.More small defect features are obtained by adding a small target de-tection layer,after which the FPN algorithm is added to fuse the feature information of deep and shallow layers to improve the feature synthesis of deep layers.The defect percentage is also amplified through image segmentation operation to improve the accuracy.The re-sults show that compared with the result of the method before optimization,the detection accuracy of PCB image is improved by 1.89%,the average detection precision of defects is improved by 1.82%,and the unnecessary detection is reduced.This provides a reference for improving the efficient detection of PCB.

关 键 词：卷积神经网络 印制电路板 缺陷检测 图像处理 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]