Patch distance based auto-encoder for industrial anomaly detection

Zeqi Ma

doi:10.1016/j.eswa.2025.126537

Return

Patch distance based auto-encoder for industrial anomaly detection

2025-04-01

PRE

Zeqi Ma

Jiaxing Li

DOI：10.1016/j.eswa.2025.126537

Original

Save

Abstract

En 中文

Industrial anomaly detection (IAD) aims to classify images and locate defective areas of anomalous samples. To perform IAD, auto-encoder is widely adopted to train on anomaly-free images and infer by calculating the distance between the input and the reconstructed output. However, in practice, auto-encoder cannot restore the anomalous region well, as it loses small spatial information, leading to the unpromising performance of detection. To overcome the above limitation, a novel approach named Patch Distance Based Auto-Encoder For Industrial Anomaly Detection (PDAE) is proposed in this study. To increase the receptive field size and robustness to small spatial deviations, patch-level features are aggregated to the reconstruction process, which can help to better differentiate and locate the anomalous areas. Furthermore, to better classify the samples, cross-patch score and cross-dimension score are incorporated in the inference stage. Moreover, to reduce the bias towards the pre-trained dataset, PDAE incorporates multi-level features extracted from convolutional neural networks and class-attention in image transformers. We conducted experiments on the MVTec-AD benchmark dataset and achieved state-of-the-art performance, with 99.7 % and 98.49% AUROC scores in anomaly detection and localization, respectively. In conclusion, extensive experimental results show that our method outperforms some state-of-the-art baseline methods on the used metrics and addresses the limitations effectively.

Keywords:

Patch

Anomaly detection

Reduce bias

Convolutional neural network

Transformer