Enhanced PCB defect detection via HSA-RTDETR on RT-DETR

Area of Science:

• Computer Vision and Artificial Intelligence
• Manufacturing and Industrial Automation
• Electronic Engineering and Quality Control

Background:

• Printed Circuit Board (PCB) defects like missing holes and shorts can cause product failure.
• Automated PCB inspection is crucial for Smart Manufacturing and Industry 4.0, but faces challenges with high-density, miniaturized components, and complex backgrounds.
• Existing machine vision and AI algorithms struggle with detecting multiscale defects, especially small ones.

Purpose of the Study:

• To propose an advanced AI-based method for automated and intelligent PCB defect detection.
• To enhance the detection of small and complex defects in PCBs.
• To improve the efficiency and accuracy of PCB defect inspection systems.

Main Methods:

• Developed a novel Hierarchical Scale-aware Attention (HSA) mechanism integrated into the RT-DETR (Real-Time Detection Transformer) framework, creating the HSA-RTDETR model.
• Introduced a new backbone network (R18-Faster-EMA) for improved model efficiency.
• Redesigned the Attention-based Intra-scale Feature Interaction (AIFI) module using cascaded group attention and implemented a Hierarchical Scale-aware Pyramid Attention Network (HS-PAN) for multi-scale feature fusion.
• Designed a new loss function to accelerate convergence and prioritize small defect detection.

Main Results:

• The HSA-RTDETR method achieved a mean average precision (mAP) of 96.9% on a PCB defect dataset, outperforming existing models.
• Demonstrated significant improvements over the original RT-DETR: 5.8% increase in precision, 7.9% in recall, and 5.4% in mAP50.
• Achieved a high inference speed of 66.2 frames per second (FPS), proving its effectiveness for real-time detection of small PCB defects.

Conclusions:

• The proposed HSA-RTDETR method offers a highly effective solution for automated PCB defect detection, particularly for small and challenging defects.
• The integration of hierarchical scale-aware attention and efficient network design enhances detection accuracy and speed.
• This approach represents a significant advancement for quality control in Smart Manufacturing environments.