SymmFlow: Unsupervised anomaly detection via symmetric normalizing flow

Abstract

Anomaly detection in industrial imaging has attracted significant research interest due to its critical applications. Recent advancements have demonstrated the potential of normalizing flows for unsupervised anomaly detection. However, conventional approaches often face challenges with the degeneracy of transformed distributions, especially in scenarios where anomaly samples are both scarce and subtle. To overcome these challenges, we propose a symmetrically structured normalizing flow model called SymmFlow. SymmFlow addresses the degeneracy of transformed distributions by maintaining the positive definiteness of the covariance matrix within multivariate Gaussian distributions. A novel two-stage training strategy is also proposed to stabilize training initially with regularization and subsequently reinforce the model’s robustness through symmetrical design. Extensive experiments on MVTec, VisA, and BTAD datasets demonstrate that the proposed SymmFlow outperforms existing methods, delivering superior detection accuracy both at the image and pixel levels. The source code is available at: https://github.com/Ace-blue/SymmFlow.