RelationAD: Industrial Vision Anomaly Detection with Relation Expression Logic Constraints

In real industrial settings, anomalies involving logical constraints are common and difficult to detect using pixel-level features. This study proposes Relational Anomaly Detection (RelationAD) based on a teacher-student model, which utilizes the Relational Feature Decoupling Loss (RFDLoss) to distinguish between semantic relational features and geometric relational features. Subsequently, the output of the student model is reconstructed using relational features that aim to capture logical anomalies through multiple relational discrepancies. Additionally, this study introduces a Sparse Relation Graph Attention (SRGAT) block to balance performance and efficiency. Finally, we filter the logical anomaly samples from the Cookie, ADFI, and VisA datasets to create the Industrial Logical Anomaly (ILA) dataset. RelationAD has been rigorously tested on the MVTec LOCO AD and ILA datasets, demonstrating competitive performance relative to existing solutions.