A Semi-Automatic Labeling Framework for PCB Defects via Deep Embeddings and Density-Aware Clustering.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Printed circuit board (PCB) inspection faces challenges with costly and slow labeling due to subtle defects and imbalanced data.
• Existing methods struggle with low-contrast defects in complex backgrounds, hindering efficient quality control.

Purpose of the Study:

• To develop a semi-automatic labeling pipeline for PCB inspection to overcome data labeling bottlenecks.
• To convert anomaly detection outputs into usable class labels, improving the efficiency of the inspection development lifecycle.

Main Methods:

• A pipeline was developed using image cropping, interchangeable embeddings (HOG, ResNet-50, ViT-B/16), and clustering (k-means, GMM, HDBSCAN).
• Cluster-level verification was performed using representative montages for quality control.
• Different embedding and clustering methods were evaluated for their effectiveness in defect classification.

Main Results:

• ResNet-50 + HDBSCAN achieved NMI ≈ 0.290 and AMI ≈ 0.283 with ~47 clusters on 9354 defects.
• ViT-B/16 + HDBSCAN showed comparable results, indicating robustness across different embedding models.
• The macro-purity exceeding micro-purity suggests efficient clustering for one-shot decisions, potentially reducing operator decisions by ~200×.

Conclusions:

• The proposed workflow offers an auditable and flexible path from anomaly localization to scalable supervision for PCB inspection.
• This approach prioritizes labeling productivity, directly addressing a key industrial bottleneck in PCB inspection development.
• The semi-automatic labeling pipeline enhances the efficiency and scalability of defect classification in automated visual inspection systems.