基于多通道特征融合学习的印制电路板小目标缺陷检测  

Small defects detection of PCB based on multi-channel feature fusion learning

作　　者：张莹[1] 邓华宣 王耀南 吴成中 吴琳[1] Zhang Ying;Deng Huaxuan;Wang Yaonan;Wu Chenzhong;Wu Lin(College of Automation and Electronic Information,Xiangtan University,Xiangtan 411105,China;National Engineering Research Center of Robot Visual Perception and Control Technology,Changsha 410082,China;Jiangxi Province Communication Terminal Industry Technology Research Institute Limited,Ji′an 343099,China)

机构地区：[1]湘潭大学自动化与电子信息学院,湘潭411105 [2]机器人视觉感知与控制技术国家工程研究中心,长沙410082 [3]江西省通讯终端产业技术研究院有限公司,吉安343099

出　　处：《仪器仪表学报》2024年第5期10-19,共10页Chinese Journal of Scientific Instrument

基　　金：江西省重大科技研发专项(20232ACC01007)资助。

摘　　要：提出了一种多通道特征融合学习的印制电路板小目标缺陷检测网络YOLOPCB,首先删除YOLOv7主干网络中最后一组MPConv层与E-ELAN层,去掉融合层的ECU模块与20×20的预测头,使用跨通道信息连接模块串联精简后的主干和融合网络;其次设计了浅层特征融合模块与新的anchors匹配策略,增加了两个低层次、高分辨率检测头;最后将YOLOv7主干网络中的3个E-ELAN作为输入,将融合层中最底部的E-ELAN和两个拼接模块作为输出,使用自适应加权跳层连接以增加同维度内信息量。在PCB Defect公开数据集上平均精度达到94.9%,检测速度达到45.6 fps;最后在企业现场制作的Self-PCB数据集中,YOLOPCB达到了最高精度76.7%,比YOLOv7检测精度提升了6.8%,能有效提高印制电路板小目标缺陷检测能力。The paper proposes a YOLOPCB network for small defects detection on printed circuit board(PCB)using multi-channel feature fusion learning.This article proposes YOLOPCB,a printed circuit board(PCB)small defects detection network that utilizes multi-channel feature fusion learning.Firstly,the last group of MPConv layer and E-ELAN layer in the YOLOv7 backbone network are removed,and the ECU module in the fusion layer and the 20×20 prediction head are eliminated.A cross-channel information connection module(CIC)is utilized to link the streamlined backbone and fusion networks.Secondly,a shallow feature fusion module(SFF)and a new anchor matching strategy are designed,which add two low-level,high-resolution detection heads.Lastly,the three E-ELAN layers in the YOLOv7 backbone network are used as inputs,while the bottommost E-ELAN and two concatenation modules in the fusion layer are used as outputs,with adaptive weighted skip-connection(AWS)to increase the information within the same dimension.The average precision on the PCB Defect datasets reaches 94.9%,with a detection speed of 45.6 fps.Furthermore,on the Self-PCB datasets obtained from on-site enterprises,YOLOPCB achieves the highest accuracy of 76.7%,which is a 6.8% improvement over the detection accuracy of YOLOv7.YOLOPCB effectively enhances the detection capability of small defects on printed circuit boards.

关 键 词：印制电路板 小目标检测 图像特征提取 多特征融合 自适应加权融合算法 

分 类 号：TH701[机械工程—仪器科学与技术] TP391.4[机械工程—精密仪器及机械]