基于YOLOv5_4layers的PCB小目标缺陷识别方法
YOLOv5_4layers Based Small Target Defect Identification Method for PCB
杨萍萍 1白艳茹1
作者信息
摘要
针对PCB表面缺陷分辨率低、小目标性以及多样性等问题,提出了一种基于YOLOv5_4layers的PCB小目标缺陷识别方法.该方法在YOLOv5 架构的基础上,通过新增采样层的方式添加小目标检测层,优化特征金字塔模型,提升小目标特征提取性能,实现小目标缺陷识别.在调整合适的锚框规格后,改进后的模型在输入640 像素×640 像素图像时,相较原模型识别精确率提升了 7.5%.在输入 736像素×736 像素图像时,识别精确率提升了1.3%,有效地提升了对PCB小目标缺陷的识别能力,对提高PCB制造过程的质量控制和产品可靠性具有实际意义.
Abstract
Aiming at the issues such as low resolution,small target size,and diversity of surface defects for PCB,a PCB small target defect identification method based on YOLOv5_4layers was proposed.Based on the YOLOv5 architecture,a small target de-tection layer was added by adding a new sampling layer and optimizing the feature pyramid model with the purpose of improving the feature extraction performance of small targets and realizing the detection of smaller targets.After adjusting the appropriate an-chor frame specifications,the improved model has a 7.5%higher detection accuracy than the original model while the image input size is 640 pixels×640 pixels.When the image input size is 736 pixels×736 pixels,the accuracy rate is increased by 1.3%,which effectively improves the identification ability of PCB small target defects,and has practical significance for improving the quality control and product reliability of PCB manufacturing process.
关键词
PCB/小目标缺陷识别/深度学习/YOLOv5_4layers/特征提取Key words
PCB/small target defect identification/deep learning/YOLOv5_4layers/feature extraction引用本文复制引用
基金项目
中央高校基本科研业务费专项(FRF-DF-22-12)
出版年
2024
NETL
NSTL
万方数据