A Novel MAE-Based Self-Supervised Anomaly Detection and Localization Method

Abstract

Despite significant advancements in self-supervised anomaly detection, multi-class anomaly detection tasks still pose substantial challenges. Most existing methods require individual network training for each category of objects. This paper presents a novel end-to-end approach for multi-class anomaly detection: self-supervised Mask-pretrained Anomaly Localization Autoencoder (MALA). Firstly, the masked autoencoder (MAE) and Pseudo Label Prediction Module (PLPM) are utilized to recover and perceive normal image patterns. Subsequently, the encoder weights are frozen for further end-to-end network training to predict anomalous maps directly. Token Balance Module(TBM) facilitates anomalous perception and improves anomaly segmentation. By utilizing the Visual Transformer and employing image inpainting as a proxy task, remarkable generalization results are achieved. The proposed method demonstrates its applicability across diverse styles of industrial products. Experiments are conducted on MVTech AD, VisA, KolektorSDD2, and MT datasets, achieving state-of-the-art results in multi-task anomaly detection and segmentation tasks. Specifically, we obtain image AUROC of 98.% and pixel AUROC of 97.1% on the MVTech AD dataset, pixel AUROC of 97.1% on the VisA dataset, and pixel AUROC of 98.7% on the KolektorSDD2 dataset.

Keywords

• Defect localization
• self-supervised learning
• visual transformer (ViT)
• masked autoencoder (MAE)
• industrial products