A unified multi‐class anomaly detection model based on reverse distillation

Abstract

Listen

Translate article

AbstractThe multi‐class single‐model detection paradigm is a prevalent design for industrial anomaly detection, exhibiting suitability for varying industrial classes and dynamic, flexible production scenarios. This approach not only enhances model adaptability but also minimizes maintenance costs. However, the current popular methods are susceptible to the ‘copying shortcut’ phenomenon, which constrains their performance on benchmark datasets. To overcome this limitation, this article proposes a multi‐class anomaly detection model: UniRD, based on the reverse distillation method. This model creates an image corruptor that expands the dataset and generates ‘normal‐corrupt’ image pairs. During the training process, their correspondence is used to optimize the reverse distillation. This process greatly exploits and releases the potential of the student decoder. Furthermore, a teacher feature adaptation module is devised to enhance the compatibility between the pre‐trained model and the anomaly detection task. This has the effect of reducing the discrepancy between teacher and student features while ensuring the consistency of normal sample features. The comprehensive evaluation results in two mainstream datasets, MVTec and VisA, and demonstrates that the proposed method exhibits improvement in all indicators compared to benchmark methods. The proposed method attains the state‐of‐the‐art, substantiating its effectiveness and advancement.