LMF-Core: A Training-Free Coreset for Industrial Anomaly Detection

Hongzhou Liu

Abstract: Automated visual inspection must detect surface and structural defects while observing only defect-free samples during setup. Train-ing-free memory-bank methods on frozen ImageNet backbones reach high accuracy, but they concatenate features from several backbone layers with equal weight and rely on an iterative greedy coreset whose memory size is not directly controllable. We address both issues with LMF-Core, a training-free framework with two statistics-only modules. Variance-Guided Fusion prunes low-information channels and weights each layer by the inverse variance of its normal features. Random-Projection Coreset replaces greedy selection with a Johnson-Lindenstrauss projection and uniform subsampling, giving a single, directly controllable memory parameter. Anomaly scores are the nearest-neighbour distance to the bank. On MVTec AD, LMF-Core attains 96.3% image-level and 96.8% pixel-level AUROC (mean over five seeds) from an 8.8 MB bank, with no training, labels, or anomalous examples, providing a compact memory footprint suitable for memory-constrained inspection hardware. Baseline figures are quoted from the original publications, whereas all reported LMF-Core results are obtained experimentally.

Keywords: Industrial anomaly detection, visual inspection, unsupervised anomaly detection, memory bank, feature fusion, random projection, computer vision