Reproduce “Supervised Anomaly Detection for Complex Industrial Images” on OpenResearch

About this paper

Aimira Baitieva, David Hurych, Victor Besnier, Olivier Bernard

Researchers at Valeo.ai developed a new dataset, VAD, featuring complex real-world industrial images with diverse defects, including logical and unseen types. They also introduced SegAD, a supervised anomaly detection method that leverages existing anomaly detectors and segmentation-based local statistics to achieve 96.5% classification AUROC on VAD, detecting both known and previously unseen defects without requiring pixel-level annotations.

Key insights

• Providing a realistic, image-level annotated industrial dataset with diverse and unseen defects, like VAD, is crucial for developing robust and practical anomaly detection methods.
• Combining anomaly maps from base detectors with local, segment-wise statistical features and a robust tree-based classifier (Boosted Random Forest) can effectively leverage weak supervision to enhance anomaly detection performance.
• Supervised anomaly detection can be made robust to unseen defect types by integrating general anomaly detection principles and training on diverse, real-world examples, thereby outperforming purely supervised classifiers.

Summary by alphaXiv.