SCMEO-DETR provides a lightweight and high efficiency solution for PCB defect detection based on RT-DETR

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Printed Circuit Board (PCB) defect detection is critical for electronic manufacturing reliability.
• Traditional Automated Optical Inspection (AOI) systems exhibit high false alarm rates.
• Existing deep learning models struggle with detecting small-scale PCB defects efficiently.

Purpose of the Study:

• To develop a lightweight and effective end-to-end detector for real-time PCB defect detection.
• To address limitations of current AOI and deep learning methods in detecting subtle defects.

Main Methods:

• Proposes SCMEO-DETR, a novel detector built upon the RT-DETR architecture.
• Integrates Sparse Enhanced Pyramid Network (SEPN) and Context-Guided Feature Pyramid Network (CGFPN) for enhanced feature extraction.
• Employs an Optimized and Adaptive Backbone (OA-backbone) and an improved Enhanced Upsampling Convolution Block with Shift Channel Mix (EUCB-SC) for efficiency.
• Utilizes an upgraded Multi-Head Multi-Scale Intra-Scale Feature-fusion Module (M2-IFM) for improved feature interaction.

Main Results:

• SCMEO-DETR achieved a mean Average Precision (mAP@0.5) of 97.10% on the PKU-Market-PCB dataset.
• Demonstrated a 5.28% improvement over RT-DETR-R18.
• Reduced model parameters by 50% and Giga Floating Point Operations (GFLOPs) by 39% compared to RT-DETR-R18.

Conclusions:

• SCMEO-DETR offers superior performance in accuracy and efficiency for PCB defect detection.
• The model's lightweight design makes it highly deployable for real-time applications in electronic manufacturing.
• This approach effectively tackles the challenges of lighting variations and small defect detection.