PCB Defect Detection via Local Detail and Global Dependency Information

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Printed circuit board (PCB) surface defects pose significant economic risks due to production environment impacts.
• Advancements in PCB manufacturing yield defects that are small and stylistically diverse, challenging traditional detection methods.
• Accurate PCB surface defect detection is crucial for maintaining production quality and minimizing losses.

Purpose of the Study:

• To propose a novel framework, Defect Detection TRansformer (DDTR), for enhanced PCB surface defect detection.
• To address the challenge of detecting small, diverse defects by integrating global and local feature extraction.
• To improve the accuracy and visualization capabilities of PCB defect detection systems.

Main Methods:

• Developed the Defect Detection TRansformer (DDTR) framework, combining Convolutional Neural Networks (CNNs) and transformer architectures.
• Employed Residual Swin Transformer (ResSwinT) in the backbone for multi-scale feature extraction, capturing both local details and global dependencies.
• Integrated spatial and channel multi-head self-attention (SCSA) in the neck for focused feature selection and utilized cascaded detectors/classifiers in the head for accuracy.

Main Results:

• Achieved the highest F1-score on the PKU-Market-PCB and DeepPCB datasets compared to existing methods.
• Produced the most informative visualization results, aiding in the understanding and interpretation of detected defects.
• Ablation experiments validated the effectiveness and individual contributions of the DDTR framework's modules.

Conclusions:

• The proposed DDTR framework significantly enhances PCB surface defect detection accuracy, especially for small and diverse defects.
• The combination of CNNs and transformers, along with SCSA, provides robust feature extraction and selection capabilities.
• DDTR offers a promising solution for improving quality control in PCB manufacturing, reducing economic losses.