Using a Small Dataset in Learning for Object Anomaly Detection

This paper introduces a novel approach to object anomaly detection using an ordered ensemble method with Procrustes distance, emphasizing data efficiency with small training sets. Unlike traditional deep autoencoder methods, which rely on pixel-wise reconstruction and require large datasets (e.g., 200 images per category in the MVTec AD dataset), our method leverages Procrustes distance to measure structural disparities between object feature shapes after translation, rotation, and scaling. By computing minimum Procrustes distances from a small set of 30 normal images per category, we derive robust thresholds for classifying objects as normal or anomalous. Evaluated on five MVTec AD categories (metal nut, cable, bottle, hazelnut, transistor), our approach achieves superior accuracy (e.g., 100% for metal nut and cable) compared to deep autoencoders, demonstrating robustness across rigid and deformable objects. This data-efficient method offers significant advantages for industrial inspection, where acquiring large defect-free datasets is challenging.