Unsupervised Visual Anomaly Detection with Score-Based Generative Model

Yapeng Teng, HaoYang LI, Fuzhen Cai, Ming Shao, Siyu Xia

Abstract: We consider leveraging the deviated outputs and gradient information from generative models due to out of distribution samples in anomaly detection (AD). Visual anomaly detection has been critical problems and widely discussed. However, in many application scenes, abnormal image samples are very rare and difficult to collect. In this paper, we focus on the unsupervised visual anomaly detection and localization task through a score-based generative model applicable to more general cases. Our work is inspired by the fact that injected noises to the original image through forward diffusion process may reveal the image defects in the reverse process (i.e., reconstruction). First, due to the differences of normal pixels between the reconstructed and original images, we propose to use a score-based generative model and associated score values as metric to gauge the defects. Second, to accelerate inference process, a novel $T$ scales approach is developed which reduces the use of redundant information from adjacent moments while leverages the information provided by the score model at different moments. These two practices allows our model to improve the generalization of AD in an unsupervised manner, but maintain a reasonable speed. We evaluate our method on several datasets to demonstrate its effectiveness.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)