首页|期刊导航|安徽大学学报(自然科学版)|基于FEMB-YOLO的工业PCB异常检测方法

基于FEMB-YOLO的工业PCB异常检测方法OA

Industrial PCB anomaly detection method based on FEMB-YOLO

中文摘要英文摘要

印刷电路板(printed circuit board,简称PCB)在现代工业中具有重要地位,PCB异常检测成为保障工业高效生产的必要手段.针对工业PCB异常检测中检测精度与效率难以平衡的问题,本研究提出了一种基于频域增强移动倒置瓶颈卷积的PCB异常检测方法,名为FEMB-YOLO(frequency-enhanced mobile inverted bottleneck convolutional YOLO).该方法以YOLOv8目标检测网络为基线,设计了频域增强移动倒置瓶颈卷积(FEMB)模块,通过通道分离机制和卷积模拟实现PCB异常区域的频域增强,在保持轻量化的同时提高模型对异常处频域纹理特征的利用能力.该模块先将输入图像特征按通道维度拆分,再分别处理PCB平滑区域和异常边缘细节纹理,并通过自适应频域融合模块进行特征对齐,以提高模型对异常区域的感知能力.在DeepPCB数据集上的实验结果表明,FEMB-YOLO方法在6种典型PCB异常类型检测中表现优异,在交并比阈值为50%时,平均精度均值mAP50达到0.948,推理每秒帧数达到322.57,证明了该方法在工业PCB异常检测中的可行性和有效性.

Printed Circuit Boards(PCBs)hold significant importance in modern industry,and PCB anomaly detection has become an essential means to ensure efficient industrial production.To address the challenge of balancing detection accuracy and efficiency in industrial PCB anomaly detection,this study proposes a PCB anomaly detection method based on Frequency-Enhanced Mobile Inverted Bottleneck Convolution(FEMB),termed FEMB-YOLO.This method utilizes the YOLOv8 object detection network as the baseline and designs a FEMB module.The module achieves frequency-domain enhancement of PCB anomalous regions through channel separation mechanisms and convolution simulation,improving the model's utilization of frequency-domain texture features at anomalous defect locations while maintaining lightweight characteristics.Specifically,the module splits input image features along the channel dimension to separately process PCB smooth regions and anomalous edge detail textures,and employs an adaptive frequency-domain fusion module for feature alignment to enhance the model's perception capability for anomalous regions.Experimental results on the DeepPCB dataset show that FEMB-YOLO achieves an mAP50 of 94.8%at an IoU threshold of 0.50 and an inference speed of 322.57 FPS,demonstrating its feasibility and effectiveness for industrial PCB defect detection.

鲍文霞;马源岑;王年

安徽大学电子信息工程学院,安徽 合肥 230601安徽大学电子信息工程学院,安徽 合肥 230601安徽大学电子信息工程学院,安徽 合肥 230601

信息技术与安全科学

PCB异常检测频域增强移动倒置瓶颈卷积轻量化

PCBanomaly detectionfrequency-domain enhancementmobile inverted bottleneck convolutionlightweight

《安徽大学学报(自然科学版)》 2026 (5)

24-31,8

福建省科技重大专项项目(2024HZ025022)合肥市自然科学基金资助项目(HZR2411)

10.3969/j.issn.1000-2162.2026.05.004

评论