Memory-Augmented Autoencoder for Industrial Anomaly Detection

Abstract

Unsupervised anomaly detection in industrial manufacturing requires identifying defective products utilizing only defect-free training images. Reconstruction-based methods using autoencoders are widely applied but suffer from limited discriminative power when defects occupy small image regions. This work introduces a Memory-Augmented Autoencoder (MAA) that combines a U-Net reconstruction component with a patch-level memory bank for nearest-neighbor-based anomaly scoring. During inference, MAA computes a hybrid anomaly score from both the reconstruction error and the memory distance to produce image-level predictions and spatial anomaly heatmaps. Evaluated on the bottle category of the MVTec AD benchmark, MAA attains an image-level AUC-ROC of 0.985 and an Average Precision of 0.996, surpassing all autoencoder-based baselines including SSIM AE (0.930) and VAE (0.910) and performing on par with the pretrained PaDiM method using a ResNet-18 backbone (0.982), while relying on only 168 defect-free training images and no ImageNet pretraining. A pixel-level AUC-ROC of 0.895 indicates a reasonable spatial localization capability. Ablation studies demonstrate that the memory-based component is the primary driver of detection performance, outperforming the reconstruction component by 16.4 percentage points in AUC-ROC, which is attributable to the robustness of patch-level memory matching for localized structural defects. As the present evaluation is confined to a single object category, the reported results are best interpreted as a proof of concept rather than as evidence of broad generalization.