Enhancing Industrial Anomaly Detection Using Edge Image Reconstruction with Neighbor Masked Convolutional Transformer Block

Abstract

Anomaly detection has become increasingly popular in recent years, where reconstruction-based approaches have been widely explored in. This approach usually assumes that the model can reconstruct normal patterns well but fails in anomalies, so that anomalies can be detected by evaluating the reconstruction error. However, it is usually difficult to control the model generalization boundaries in practice, because models with too much generalization ability will reconstruct the anomalous regions well, making them indistinguishable, and models with poor generalization ability fail to reconstruct the normal regions. In order to solve the above problems, we propose a new network that will reconstruct the original RGB image from edges of the image grayscale values by fusing the improved Neighbor Masked Convolutional Transformer Block, which achieves new and excellent results on the industrial anomaly detection MVTec-AD dataset.