|国家科技期刊平台
首页|期刊导航|红外技术|基于PA-YOLO v5的印制电路板缺陷检测

基于PA-YOLO v5的印制电路板缺陷检测OA北大核心CSTPCD

PCB Defect Detection Based on PA-YOLO v5

中文摘要英文摘要

针对印制电路板裸板布局复杂,在对其表面进行缺陷检测时存在被检测图像对比度不高、亮度不均匀、缺陷位置小、形状不规则等特点,在增加网络深度时会造成参数量大、出现过拟合现象、丢失部分特征信息等问题,提出了基于YOLO v5 与混合注意力机制融合,精度更高的印制电路板检测模型PA-YOLO v5(precision and attention-YOLO v5),抑制一般特征的干扰,保证网络提取特征时更加关注缺陷目标细节特征.并引用自适应双向特征融合模块(BiFPN)网络,对每个特征图的尺度不同进行充分利用,对不同的检测目标赋予不同权重,提高网络的各个特征表达能力,最后利用FReLU激活函数,通过将ReLU增加空间拓展成为一个 2D激活函数,增强感受野对细节捕捉的能力,提高模型的鲁棒性和泛化性.在DeepPCB数据集中对 6 种缺陷分别进行测试,实验结果表明,文中提出的PA-YOLO v5的检测模型在该数据集上的准确率可达 99.4%,并同时设置消融实验和对比实验验证了该模型的有效性.

The bare complex layout of PCBs cause low contrast,uneven brightness,small defect positions,and irregular shapes in detected images,resulting in a large number of parameters,overfitting,and loss of feature information with increasing network depth.In this study,a PCB detection model PA-YOLO v5 based on YOLO v5 and mixed attention mechanism fusion with higher accuracy is proposed to suppress interference from general features and ensure that the network pays more attention to the detailed features of defect targets during feature extraction.The adaptive bidirectional feature pyramid network(BiFPN)is taken as reference to fully utilize the different scales of each feature map,thereby assigning different weights to different detection targets,to improve the network's ability to express various features.Finally,the FReLU activation function is used to expand the ReLU space into a 2D activation function,which enhances the receptive field's ability to capture details and improves model robustness and generalization.Six types of defects were tested using the DeepPCB dataset,and the experimental results showed that the proposed PA-YOLO v5 detection model achieved an accuracy of 99.4%.The effectiveness of the model was verified through ablation and comparative experiments.

陈锦妮;拜晓桦;李云红;田谷丰

西安工程大学 电子信息学院,陕西 西安 710048

计算机与自动化

缺陷检测注意力机制印刷电路板深度学习

defect detectionattention mechanismprinted circuit boarddeep learning

《红外技术》 2024 (006)

654-662 / 9

陕西省科技计划项目(2022GY-053).

评论