Meta-ensemble of fastFlow and PaDiM for efficient industrial anomaly detection

Abstract

Accurate detection of defects is critical in industrial manufacturing to ensure optimum product quality and minimize losses. Anomaly detection models help identify unexpected deviations from normal product features. However, those present approaches are limited by the reliance on a single model, which restricts learning rare and evolving defect patterns. Therefore, in this work, we present a meta-ensemble anomaly detection framework named FlowDiMBoost, which utilizes FastFlow and PaDiM, backed by ResNet-18, on the MVTec AD anomaly detection dataset. FastFlow uses normalizing flows over pretrained embeddings to model the distribution of normal samples, whereas PaDiM uses patch distribution modeling. The outputs of those two heterogeneous models are then fused using an XGBoost meta-learner, enhancing detection performance. The models are trained using only the normal samples, allowing real-world defect detection without prior exposure to anomalies. The experimental results demonstrate that FlowDiMBoost yields better performance than individual models, achieving AUROC gains of 0.52% to 7.24% and F1-score gains of 0.55% to 11.21% by delivering consistent improvements across a variety of types of MVTec AD object categories.