Imitating Anomaly-Guided Student-Teacher Model for Unsupervised Anomaly Detection
Qing Zhao, Yan Wang, Yuxuan Cosmi Lin, Shaoqi Yan, Boyang Wang, Wei He, Xuening Wu, Yang Chang, Wenqiang Zhang · 2023
Defect detection in industrial settings encounters challenges such as limited defect samples and the availability of only normal samples. Existing defect detection methods primarily fall into two categories: feature embedding-based and reconstruction-based. Knowledge distillation, a representative feature-embedding-based approach, has gained prominence due to its lightweight nature, efficiency, and ease of deployment. Nevertheless, conventional knowledge distillation networks compromise defect detection accuracy due to structural and input stream similarities. In response, we introduce IAG, an unsupervised anomaly detection model employing mutual constraints and a novel hybrid noise generation strategy to simulate anomalous structures and textures. Firstly, to generate anomalies with structures and textures more akin to real defects, we devise a hybrid noise generation module based on the structural edge shapes of withered-leaf noise, statistical noise, and the texture of tiling noise. Secondly, we design a mutual constraint framework that individually constrains a single noise texture, enhancing the model's ability to recognize anomaly features. Thirdly, we propose a novel evaluation metric for small defect detection, addressing the limitation of existing metrics in effectively evaluating model performance for small defect detection. Through comprehensive ablation experiments, we substantiate the efficacy of the proposed model in anomaly detection and localization.