An end-to-end framework for hyperspectral detection of automotive paint anomalies

Abstract

This study introduces a system-level framework that defines a structured pipeline for unsupervised car-paint defect detection using Hyperspectral Imaging (HSI), marking a novel application in automotive quality inspection. The proposed end-to-end method includes black-and-white calibration, Savitzky-Golay filtering, Minimum Noise Fraction (MNF) transformation, multiple anomaly detection techniques (Isolation Forest, Robust Random Cut Forest, One-Class SVM, Reed-Xiao Detector, Autoencoder), and 2D-Total Variation (2D-TV) for post-processing. Experimental results on two collected datasets (Megane and Skoda) show that RRC consistently achieves the top scores across ranking- and threshold-based metrics. The 2D-TV method enhances AUC scores by 2-4% for the Megane dataset and 10-15% for Skoda by reducing noise and preserving structural details. This work demonstrates the feasibility and effectiveness of HSI for unsupervised paint defect detection, advancing automotive inspection technologies.