MDEB-YOLO: A Lightweight Multi-Scale Attention Network for Micro-Defect Detection on Printed Circuit Boards

Area of Science:

• Computer Vision
• Machine Learning
• Electronics Manufacturing

Background:

• Printed Circuit Board (PCB) defect detection is vital for electronics quality control.
• Existing deep learning models struggle with small, irregular defects and background noise, balancing accuracy and speed.
• Challenges include intricate circuitry, minute defect scales, and varied morphologies.

Purpose of the Study:

• To develop a lightweight, real-time detection network (MDEB-YOLO) for PCB micro-defects.
• To enhance the model's ability to perceive subtle geometric variations and extract irregular defect features.
• To improve the representation of small targets and reduce computational complexity for industrial applications.

Main Methods:

• Proposed the Efficient Multi-scale Deformable Attention (EMDA) module for enhanced feature extraction.
• Introduced a Bidirectional Residual Multi-scale Feature Pyramid Network (BRM-FPN) to mitigate feature loss.
• Developed a Lightweight Grouped Convolution Head (LGC-Head) to reduce model size and complexity.

Main Results:

• MDEB-YOLO achieved 95.9% mAP and 80.6 FPS on the PKU-Market-PCB dataset.
• Demonstrated a 1.5% mAP improvement and 26.5% faster inference compared to baseline models.
• Significantly improved detection accuracy for mouse bite (3.7%) and spur (4.0%) defects.

Conclusions:

• MDEB-YOLO offers superior accuracy and real-time performance for PCB micro-defect detection.
• The proposed EMDA, BRM-FPN, and LGC-Head modules effectively address challenges in detecting small, irregular defects.
• The model holds significant value for industrial electronics manufacturing quality control.