Unsupervised Visual Anomaly Detection Using Self-Supervised Pre-Trained Transformer

In the various industrial manufacturing processes, the automatic visual inspection system is an essential part as it reduces the chances of delivering defective products and the cost of training and hiring experts for manual inspection. In this work, we propose a new unsupervised anomaly detection method inspired by the masked language model for the automatic visual inspection system. The proposed method consists of an image tokenizer and two subnetworks, a reconstruction subnetwork, and a segmentation subnetwork. We adopt a pre-trained self-supervised vision Transformer model to use it as an image tokenizer. Our first subnetwork is trained to predict the anomaly-free patch tokens and the second subnetwork is trained to produce anomaly segmentation results from both the reconstructed and input patch tokens. During training, only the two subnetworks are optimized, and parameters of an image tokenizer are frozen. Experimental results show that the proposed method exhibits better performance than conventional methods in detecting defective products by achieving 99.05% I-AUROC on MVTecAD dataset and 94.8% I-AUROC on BTAD.