.png){kind=icon}
.png){kind=icon}
Return
YOLO-MDMN: a new lightweight algorithm for efficient PCB defect detection
.png){kind=icon}
PRE
AI
DOI：10.1088/1361-6501/ae46b6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Printed circuit boards (PCBs) defect detection constitutes a critical stage in electronic manufacturing, serving as a vital safeguard for product reliability, functional safety, and overall quality in industries such as automotive, aerospace, and consumer electronics. However, existing methods still face some key limitations, including the frequent omission of micron-scale defects caused by insufficient receptive fields and coarse feature representations, alongside high computational costs that hinder real-time deployment on resource-constrained production equipment. Therefore, this paper proposes You Only Look Once (YOLO)-MDMN—an enhanced variant of YOLOv12n. First, the improved Monte Carlo receptive field module is integrated into the A2C2f module, which utilizes stochastic multi-scale sampling to reduce both target fragmentation and computational overhead. Second, the dynamic spatial-preserving feature neck is utilized to combine Space-to-Depth encoding with content-aware upsampling and adaptive fusion to preserve high-frequency details during scaling. Third, the Detect_MBConv detection head incorporates adaptive expansion ratios, scale-aware depthwise convolution, and dual-pathway attention for improved efficiency. Finally, the Shape-Nwd loss function is applied to combine geometric constraints with adaptive Gaussian modeling for stable micro-defect localization. Experiments on PKU-Market-PCB dataset show YOLO-MDMN achieves 93.8% mAP50, outperforming YOLOv12n by 3.5%, while reducing parameters by 39.7% and computation by 40.6% and achieving real-time inference at 344.8 FPS, which demonstrates efficacy for resource-constrained industrial deployment.
Keywords:
PCB defect detection
YOLOv12n
Monte Carlo receptive field
lightweight model
real-time inference
Journal
IF:
3.4
Papers:
2.6K
Citations:
2.3W