Anomaly Detection in Industrial Images Using Normalizing Flow with Self-Attention Mechanisms

Abstract

Anomalies in industrial images can be categorized into logical anomalies and structural anomalies. Logical anomalies refer to irregularities such as object deficiency, excess, or misplacement, while structural anomalies indicate impurities such as dirt, scratches, or foreign matter inclusion. In conventional anomaly detection methods based on normalized flow, variable transformation is performed considering local information in the feature map. While these methods generally have good detection performance with respect to structural anomalies, they are not good at detecting logical anomalies. In this study, to address this issue, we propose the Multi-head Self Attention Flow (MSAFlow), which introduces a self-attention mechanism into the normalization flow to capture the relationships between features during variable transformation. The proposed method was evaluated in comparison with the conventional convolutional layer-based normalization flow on the MVTecLOCO dataset, achieving a 5% improvement in the average AUROC across all categories.