摘要
本文针对印制电路板(Printed Circuit Board,PCB)的缺陷检测问题,研究一种基于改进YOLOv5的PCB缺陷检测模型,构建常见PCB缺陷图像和背景图像的训练数据集,以及缺陷图像和无缺陷图像的测试数据集。为提高YOLOv5模型全局特征捕获能力,在CSP模块的ResNet中,融入Transformer的多头注意力机制,构建改进后的YOLOv5网络结构。结果表明:改进后的模型更适合PCB缺陷的检测,对非缺陷图像的检测精度提高了11.40%。
Aiming at the problem of Printed Circuit Board(PCB)defect detection,this paper studies a PCB defect detection model based on improved YOLOv5,and constructs training data sets of common PCB defect images and background images,as well as test data sets of defect images and defect-free images.In order to improve the global feature capture capability of YOLOv5 model,the multi-head attention mechanism of Transformer is integrated into ResNet of CSP module to build the improved YOLOv5 network structure.The results show that the improved model is more suitable for PCB defect detection,and the detection accuracy of non-defect images is increased by 11.40%.
作者
徐海达
常正方
罗轶
李彦龙
张健滔
Xu Haida;Chang Zhengfang;Luo Yi;Li Yanlong;Zhang Jiantao
出处
《计量与测试技术》
2023年第11期30-32,共3页
Metrology & Measurement Technique
基金
国家自然科学基金资助项目(项目编号:No.52175102)。