Unsupervised Defect Detection in Industrial Images with Conformal Anomaly Detection and VAEs
e-Journal of Nondestructive Testing · 2026-07 · 31 (7)
DOI: 10.58286/33081 ↗Kurz gefasst
Fachartikel in der Zeitschrift e-Journal of Nondestructive Testing, 2026-07. Verfasst von Flora Estermann und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Anomaly detection plays a major role in industrial quality control and predictive maintenance, where early identification of defects prevents costly downtime and ensures operational safety. However, supervised approaches for automatic anomaly detection require large annotated datasets of defective samples, which are expensive to collect and rarely representative of all possible anomalies. In this work, we propose an unsupervised framework combining a Variational Autoencoder (VAE) and Conformal Anomaly Detection (CAD) to detect defects from normal data, with CAD providing statistically calibrated p-values for uncertainty estimation. Results show that the proposed VAE+CAD pipeline improves sensitivity by up to +12% over standard reconstruction-based methods while maintaining high specificity on sufficiently large datasets. However, when the calibration set is too small, the estimated normal distribution becomes unreliable, leading to degraded specificity. These findings highlight both the potential and the limitations of CAD for unsupervised defect detection, representing, to our knowledge, the first evaluation of CAD on real industrial image data.
Zitieren
APA
Estermann, F., Guy, P., Delachartre, P., & Kaftandjian, V. (2026). Unsupervised Defect Detection in Industrial Images with Conformal Anomaly Detection and VAEs. e-Journal of Nondestructive Testing, 31(7). https://doi.org/10.58286/33081
Vancouver
Estermann F, Guy P, Delachartre P, Kaftandjian V. Unsupervised Defect Detection in Industrial Images with Conformal Anomaly Detection and VAEs. e-Journal of Nondestructive Testing. 2026;31(7). doi:10.58286/33081
BibTeX
@article{Estermann2026,
author = {Estermann, Flora and Guy, Philippe and Delachartre, Philippe and Kaftandjian, Valerie},
title = {Unsupervised Defect Detection in Industrial Images with Conformal Anomaly Detection and VAEs},
journal = {e-Journal of Nondestructive Testing},
year = {2026},
volume = {31},
number = {7},
doi = {10.58286/33081}
}
EUVIMED