Semantically Guided Counterfactual Model for Multiclass Anomaly Detection

Reconstruction-based anomaly detection is appealing for industrial inspection because it reconstructs anomaly-free references and produces interpretable pixel-level residual maps. However, in multiclass settings, it often suffers from, first, identity mapping, where abnormal regions are overreconstructed and residuals vanish, and second, cross-category feature entanglement, which introduces artifacts and weakens localization. Residual scoring is further biased because it mixes true defect discre