MFE-YOLO: A Multi-Scale Feature Enhanced Network for PCB Defect Detection with Cross-Group Attention and FIoU Loss.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Automated quality control in manufacturing relies on accurate defect detection in Printed Circuit Boards (PCBs).
• Existing deep learning models like YOLOv5 often exhibit overconfident errors due to a lack of predictive uncertainty quantification, leading to false alarms in industrial inspection.
• Probabilistic calibration is essential for trustworthy automated inspection systems.

Purpose of the Study:

• To propose MFE-YOLO, a Bayesian-enhanced detection framework built upon YOLOv5.
• To improve both the accuracy and operational reliability of PCB defect detection by integrating uncertainty-aware mechanisms.
• To address the challenges of small targets and complex backgrounds in industrial PCB inspection.

Main Methods:

• Constructed a multi-background PCB defect dataset to improve model generalization.
• Integrated the Convolutional Block Attention Module (CBAM) as a feature-wise uncertainty weighting mechanism.
• Developed a novel probabilistic FIoU loss function for improved bounding box regression and localization uncertainty capture.

Main Results:

• MFE-YOLO achieved state-of-the-art performance with mAP@0.5 of 93.9% and mAP@0.5:0.95 of 59.6%, outperforming YOLOv8 and EfficientDet.
• The framework demonstrated significantly better-calibrated confidence scores, reducing false alarms.
• Achieved improved detection accuracy and reliability for small defects in complex backgrounds.

Conclusions:

• MFE-YOLO offers a deployable, uncertainty-aware solution for high-throughput PCB inspection.
• The proposed Bayesian enhancement leads to more trustworthy automated quality control in manufacturing.
• This work advances the reliability of automated visual inspection systems through uncertainty quantification.