A Lightweight Denoising Student-Teacher Network for Anomaly Detection and Localization

Anomaly detection and localization are important tasks in the field of computer vision. Currently, most anomaly detection methods focus on the effectiveness of detection and localization, while neglecting their practical applications. This paper proposes a lightweight denoising student-teacher network for anomaly detection and localization that reduces model parameters without compromising the effectiveness of detection and localization. The proposed network utilizes a pretrained teacher network, a lightweight denoising student network, and an attention-guided segmentation network, trained under the supervision of synthetic anomaly masks. Experiments on the MVTec AD industrial inspection benchmark dataset demonstrate that the proposed network not only reduces model complexity but also improves the effectiveness of anomaly detection and localization.