Reverse Distillation Support Vector Data Description for Unsupervised Anomaly Detection
Zhipeng Liu, Gang Wu, Jianmin Lan · 2023
Anomaly detection is an essential aspect of industrial production, but it faces the difficulty of limited samples of anomalous data and too many flawed styles. Unsupervised learning is able to understand the data structure from normal samples and is able to deal with such issues better. The reverse distillation strategy, which builds a teacher encoder and a student decoder to enable the student network to learn the feature representation of the teacher network using a self-encoding mechanism, performs very well in this study. This paper presents an enhanced version of the reverse distillation model by integrating the coordinate attention mechanism and support vector data description. During the training process, we construct a hypersphere space to serve as a representation of the implicit space of the model. Subsequently, we aim to enhance the discernment of abnormal feature samples by decreasing the distance between the features of normal samples inside the implicit space. Furthermore, the optimization of the objective function of context similarity is employed to mitigate overfitting issues. As a result, the proposed reverse distillation approach achieves superior performance on Mvtec, with AUROC image-level score of 99.13% and AUROC pixel-level score of 98.1%.