BPGG: Bidirectional Prototype Generation and Guidance Network for Few-Shot Anomaly Localization

Few-shot anomaly localization task is pixel-level detection of unseen images with only a tiny amount of anomaly training samples. Bound by reality, most conventional training data are defect-free, and models are difficult to accommodate various anomaly types. To this end, we propose a bidirectional prototype generation and guidance network (BPGG), which implements non-parametric metric learning with the help of prototypes. We first trade the position of the support set and query set to construct the adaptive reverse branch. The bidirectional branch structure forces the support set and query set to align with each other and build a consistent metric space. For leveraging the benefits of regular data, we also insert the normal images into the support set and balance the proportion of normal and defective samples. Our experimental study on the MVTec anomaly detection dataset demonstrates that our proposed algorithm outperforms current few-shot SOTA methods, comparable to other unsupervised and self-supervised algorithms. Besides, our BPGG Network is general to detect various types of real-world defects and perform stable detection. The rational utilization of data and innovative architecture in our study provide a novel breakthrough for the task of anomaly location.