Discrepancy Constraint Learning for Industrial Unsupervised Anomaly Detection

Abstract

Anomaly detection plays an important role in industrial production. In this area, existing reconstruction-based methods aim to reduce the reconstruction discrepancy on the normal samples and then the abnormal samples can not be reconstructed well. However, there is a risk that the model has no pathway to know the abnormal features. Thus, these methods cannot guarantee the high reconstruction discrepancy of the abnormal samples since the anomaly is open-set. Therefore, we propose a discrepancy constraint learning method (DiCo) for the unsupervised anomaly detection, including discrepancy construction and discrepancy augmentation. First, in the discrepancy construction, we design a dual-branch input autoencoder. For the same images in each batch, one pre-trained encoder branch is fed into the normal images and another is fed into the masked images. The hierarchical reconstruction error is calculated between the masked images reconstruction output and the normal images input. Second, in discrepancy augmentation, we propose a learnable dynamic memory module to store the common normal features and reweight the abnormal features. Meanwhile, a novel anomaly score is proposed to select the larger pixel predicted values adaptively as the anomaly score to effectively improve the discrepancy between normal and abnormal samples. The experimental results illustrate that DiCo outperforms the SOTA methods on the well-known dataset MVTec with 99.5%AUC in image level and 98.8%AUC in pixel level.

Keywords: Anomaly DetectionSelf-supervised LearningDynamic MemoryAnomaly Score