Image alignment-based patch distribution framework for anomaly detection

Surface quality inspection is a crucial step in industrial manufacturing. However, it is challenging to collect an adequate quantity of abnormal samples in practice. Supervised methods require sample annotation, which is costly, so unsupervised methods that are high-speed and low-cost are more suitable for industrial applications. Among the current unsupervised methods, embedding-similarity methods haven shown excellent performance, but most of them do not preprocess the images and directly use convolutional neural networks to extract image features. While in actual scenarios, image can have some degree of offset or rotation due to machine variations. Therefore, this paper proposes a new patch distribution framework for anomaly detection, specifically a novel image alignment module is proposed to enhance the utility of the model. Image alignment reduces the dense distance between pixels during training, enabling more precise learning of the feature distribution of normal samples and reduce false positives during testing. In addition, in the feature extraction stage, the middle layer of the network is selected to extract features and establish embedding connections. This not only enhances the model's precision but also reduces its memory requirements. Experiments on the publicly available datasets MVTec AD and BeanTech AD show that our proposed new framework achieves better performance than other baseline models.