Lightweight One-Class Industrial Anomaly Detection Using a Normal-Calibrated Random-Subspace Ensemble

Abstract

Industrial anomaly detection is commonly performed under one-class conditions because defective samples are scarce or unavailable during training. This study presents a lightweight one-class anomaly detection framework that reduces the dependence of compact deep representations on a single arbitrarily selected feature subset. Intermediate features extracted from a frozen MobileNetV2 backbone are divided into five independently sampled 100-channel subspaces. A position-specific diagonal Gaussian model is fitted separately to each subspace using only normal training images. The resulting anomaly scores are robustly normalized and combined through a normal-calibrated reliability-weighting strategy, while image-level decisions are obtained from the highest-scoring local patches. To prevent information leakage, operating thresholds are determined through five-fold cross-fitted calibration using only normal training scores. The proposed compact random-subspace ensemble was evaluated on all 15 categories of MVTec AD and externally validated on the 12-category VisA benchmark. On MVTec AD, the ensemble increased the macro ROC-AUC from 0.9350 to 0.9437 and the macro F1 score from 0.8366 to 0.8504 compared with a single random 100-channel subspace. External validation on VisA showed the same improvement trend, with higher macro ROC-AUC, AP, and F1 values than the single-subspace configuration. The results demonstrate that combining multiple compact normal-data subspaces improves discrimination and reduces reliance on the evaluated single random representation while retaining a lightweight, anomaly-label-free training procedure.