Feature Configuration Optimization of PaDiM for Industrial Anomaly Detection: An Experimental Study on Eight MVTec AD Categories

Industrial anomaly detection is a critical component of machine-vision-based intelligent manufacturing, especially when defective samples are scarce and pixel-level annotations are expensive to obtain. PaDiM is a widely used unsupervised anomaly detection method that models patch-level feature distributions extracted from pretrained convolutional neural networks; its practical performance depends on the selected feature layers and the dimensionality of patch embeddings. This study proposes FO-PaDiM, a feature configuration optimization framework built on the PaDiM pipeline that keeps the original statistical modeling unchanged while exhaustively searching feature-layer combinations and embedding dimensionality. Unlike a previous exploratory version of this study, all configuration selection is performed under a leakage-free development protocol—a deterministic 80/20 split of training normals combined with multi-type synthetic anomalies—so that the MVTec AD test set is used exactly once, with a locked configuration and frozen thresholds. On eight MVTec AD categories (bottle, cable, capsule, hazelnut, metal_nut, pill, screw, transistor) with 14 configurations per category (seven non-empty ResNet-18 layer combinations × n_features in {32, 64}), the selection is found to be category-dependent: single-stage configurations (layer2, layer3 or layer1) are preferred in six categories and multi-stage combinations in two, while the previously reported universal preference for the three-layer combination does not replicate once test-set leakage is removed. Under the four evaluated weighting schemes, the layer combination remains unchanged for 62.5% of categories and n_features for 87.5%. Final locked test performance reaches a mean Image AUROC of 0.8676, Image F1 of 0.8079, Pixel AUROC of 0.9644, and Pixel F1 of 0.4582. The results provide evidence that multi-level feature configurations can be beneficial but that their advantage is category-dependent and small in magnitude, and they confirm the persistent difficulty of threshold-dependent binary localization (Pixel F1) despite strong ranking quality (Pixel AUROC).