Local and Global Context-Enhanced Lightweight CenterNet for PCB Surface Defect Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Printed circuit board (PCB) surface defect detection is critical for manufacturing quality.
• Existing computer vision methods for PCB defect detection are computationally intensive and inefficient.
• High-resolution imaging is available, but efficient analysis remains a challenge.

Purpose of the Study:

• To develop a lightweight and efficient model for real-time PCB surface defect detection.
• To improve the accuracy and inference speed of automated optical inspection (AOI) systems.
• To address the computational limitations of current computer vision approaches in PCB manufacturing.

Main Methods:

• Proposed a local and global context-enhanced lightweight CenterNet (LGCL-CenterNet).
• Introduced a two-branch lightweight vision transformer module (LGT) with local (coordinate attention) and global (Bi-Level Routing Attention) attention mechanisms.
• Incorporated a Path Aggregation Network (PANet) for feature fusion and a lightweight prediction head using depthwise separable convolutions.

Main Results:

• LGCL-CenterNet achieved improved mAP@0.5 by 2% and 1.4% compared to CenterNet-ResNet18 and YOLOv8s, respectively.
• The model requires fewer parameters (0.542M) than existing techniques.
• Demonstrated superior detection accuracy and inference speed.

Conclusions:

• LGCL-CenterNet offers enhanced real-time performance and robustness for PCB surface defect detection.
• The proposed method effectively balances detection accuracy and computational efficiency.
• This approach contributes to more efficient and reliable automated inspection in PCB manufacturing.