Spectral learning-based artificial vision for quality control

Artificial intelligence (AI) is increasingly employed in quality control (QC), yet acquiring sufficient and diverse training data in industrial environments remains a persistent challenge. To address this, we propose a novel approach using synthetic multispectral metrology images to efficiently train classification models for integration into in-line artificial vision systems. By exploiting material-specific spectral signatures, we constructed a comprehensive dataset comprising both real and physically accurate synthetic images that incorporate essential material properties. We trained and optimized CNN, SVM, KNN, and Random Forest classifiers using grid search and cross-validation. Performance evaluation based on precision, recall, and ROC AUC demonstrated that models trained on synthetic data consistently outperformed those trained solely on real images. These findings underscore the potential of spectral imaging and physics-based synthetic data to enhance automated material classification in industrial QC systems.