ResNet-SE-CBAM Siamese Networks for Few-Shot and Imbalanced PCB Defect Classification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Mass production defect detection faces challenges with small, imbalanced datasets, limiting traditional deep learning.
• Few-shot learning is crucial for adapting models to real-world industrial scenarios with minimal data.

Purpose of the Study:

• To develop and evaluate a few-shot learning model for effective product defect detection using limited data.
• To enhance model generalization, stability, and applicability in industrial settings.

Main Methods:

• Proposed a ResNet-SE-CBAM Siamese network for feature extraction, incorporating attention mechanisms and metric learning.
• Utilized triplet loss for embedding learning and a Structural Similarity Index Measure (SSIM) for sample selection.
• Implemented a high defect rate training strategy and a K-Nearest Neighbor (KNN) classifier for improved stability and reduced false negatives.

Main Results:

• Achieved 94% classification accuracy and 2% False Negative Rate (FNR) with a 20:40 good-to-defect ratio.
• Reached zero false negatives (FNR = 0%) when the number of defective samples increased to 80.
• Outperformed traditional deep learning models like YOLO in accuracy and miss rates.

Conclusions:

• The proposed metric learning approach demonstrates superior performance in few-shot defect detection compared to traditional deep learning models.
• The system offers high reliability and potential for industrial deployment, effectively addressing challenges of limited and imbalanced datasets.