Industrial Image Anomaly Detection Method Based on Improved MAE

This study explores an improved Masked Autoencoder (MAE) model for anomaly detection in common industrial scene images. Industrial image anomaly detection, as an important research topic in computer vision, can detect abnormal data deviations from normal expected behavior and ensure the normal operation of various systems. In actual industrial scenarios, the scarcity of abnormal samples, the cost of labeled data, and the lack of prior knowledge about anomalies make unsupervised learning methods widely used in the field of image anomaly detection. However, most unsupervised learning methods currently require the application of large-scale datasets. In order to make unsupervised learning applicable to small-scale industrial image datasets, we replaced the backbone network with Swin Transformer based on the Masked Autoencoder method, while improving its mask strategy and optimizer. The effectiveness and advantage of the proposed method are demonstrated through an experimental comparison with existing model, the results show that this method is better than MAE model pretrained on ImageNet, and outperforms other unsupervised learning models in multiple categories.