基于增强小目标特征提取的PCB板缺陷检测模型  

PCB Board Defect Detection ModelBased on Enhanced Small-target Feature Extraction

作　　者：季堂煜 赵倩 余文涛 梁爽 赵琰 JI Tang-yu;ZHAO Qian;YU Wen-tao;LIANG Shuang;ZHAO Yan(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海201306

出　　处：《仪表技术与传感器》2023年第4期87-92,共6页Instrument Technique and Sensor

基　　金：国家自然科学基金资助项目(61802250)。

摘　　要：针对印制电路板(PCB)表面缺陷所具有的分辨率低、小目标性以及多样性等问题,提出基于YOLOv5的增强小目标特征提取的PCB板缺陷检测模型——SPDYOLOv5模型。在主干网络引入SPDConv,提高主干网络对各尺度特征的提取能力。在主干网络最深层加入CA注意力,加强深层信息的传递能力。提出T3Head特征融合结构,在上下采样阶段融入CBAM注意力机制,加强各尺度间的信息传递能力;借助转置卷积和空间深度卷积,优化特征融合结构对小目标特征的表达能力。在训练过程中,迁移VOC预训练权重加速收敛。采用EIOU-NMS进行后处理,改善模型检测效果。实验结果表明:文中模型在北京大学开源PCB板缺陷数据集上mAP0.5可达92.4%,性能优于其他检测方法。For the problems of low resolution,small target and diversity of printed circuit board(PCB)surface defects,a PCB board defect detection model that is SPDYOLOv5 model based on YOLOv5 enhanced small target feature extraction was proposed.SPDConv was introduced into the backbone network to improve the feature extraction of each scale by the backbone network.The CA attention was added to the deepest layer of the backbone network to enhance the transmission ability of deep information.The T3Head feature fusion structure was proposed,and the CBAM attention mechanism was integrated in the up and down sampling stage to enhance the information transfer ability between scales.With the help of transposed convolution and SPDConv,the feature fusion structure s ability to express small target features was optimized.During training,migrating VOC pretrained weights accelerated convergence.EIOU-NMS was used for post-processing to improve the model detection effect.The experimental results show that the model in this paper can achieve a mAP0.5 of 92.4%on the open source PCB board defect data set of Peking University,and its performance is better than other detection methods.

关 键 词：深度学习 PCB表面缺陷检测 YOLOv5 小目标检测 迁移学习 EIOU-NMS 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]