Image-level Synthesis and Perturbation for Self-supervised Anomaly Detection

Synthetic anomaly samples can play a crucial role in unsupervised image anomaly detection. Current image anomaly synthesis methods exploit image-level operations such as overlays or embedding-level operations such as perturbation by Gaussian noise. However, they have limitations in terms of the diversity and realism of the synthesized samples, respectively. In this paper, we integrate these techniques by guiding the impact of embedding-level perturbation towards specific regions of the synthesized images. We control the impact of the embedding-level perturbation on the normal regions of the synthetic anomalies using gradient descent, in order to maintain its consistency with the normality. Training feature extractor and localization networks with more diversity over the synthesized regions contributes to enhanced robustness. The empirical results show the advantage of our approach over the state-of-the-art method.