Anomaly Detection Model Based Visual Inspection Method for PCB Board Manufacturing Process

We developed a visual inspection method for PCB board using an anomaly detection model. To improve feature extraction performance, we developed and optimized the feature extractor by comparing three types of backbone models. Then we compared two anomaly detection models with developed feature extractor as a backbone for visual inspection. Finally, we found the optimized loss function named mean-shifted contrastive loss which showed the highest accuracy in our experiment.