Revisiting segmentation-guided denoising student–teacher in anomaly detection

Anomaly detection is a critical issue that needs to be addressed in large-scale industrial manufacturing. DeSTSeg integrates a pre-trained teacher network, a denoising student encoder–decoder, and a segmentation network into a single framework. However, this network faces two main challenges. Firstly, the supervision between the teacher and student networks is only focused on the same layers, lacking additional efficient information exchange, making it difficult for the student to effectively learn knowledge. Secondly, simple image-level perturbations struggle to cope with complex industrial scenes. To address these issues, we propose a high-performance anomaly detection network that combines knowledge review and feature-level perturbation, named KR-FP. To facilitate more efficient learning from the teacher network by the student, we introduce a knowledge review (KR) module to enhance the learning pathways of the teacher–student networks. This module comprises an attention-based fusion (ABF) module and a hierarchical context loss (HCL) module. The ABF module generates multi-scale fused features by integrating features from different levels. Then, the HCL module transfers knowledge into different levels of contextual information for interaction, thereby enhancing learning outcomes. Additionally, due to the limitations of image-level data augmentation methods in complex industrial environments, we introduce a Feature Perturbation (FP) module to enhance the model’s ability to adapt to diverse and complex industrial scenes, significantly improving the model’s stability and robustness. Experiments on the MVTec AD dataset have proven the effectiveness of our approach, achieving competitive results with a pixel-level average accuracy of 78.0% and an instance-level average accuracy of 78.8%. Source code and pretrained models are available at https://github.com/chongyouxiaolu/KR-FP.