Siamese Contrastive Reverse Distillation for Industrial Anomaly Localization

The unsupervised anomaly localization based on feature distillation has demonstrated outstanding performance in industrial anomaly localization. It relies on the feature discrepancies between the student network and the teacher network to achieve anomaly localization. In these methods, the student network exclusively learns from the normal features of the teacher network during training, overlooking the explicit constraining of prior anomaly knowledge. This results in uncertainty in the feature discrepancy between the student and teacher for abnormal inputs, leading to a decrease in prediction accuracy. To inject prior anomaly knowledge into the student network during training, this paper proposes Fine-Grained CutPaste (FG-CutPaste) data augmentation strategy and Siamese Contrastive Reverse Distillation Network (SCRD). FG-CutPaste provides pseudo-abnormal samples and corresponding pixel-level pseudo-abnormal labels during the training phase of SCRD. SCRD introduces the Siamese network paradigm along with Contrastive Distillation (CD) loss. The CD loss, utilizing pseudo-abnormal samples and labels, not only reduces the discrepancy of normal features between the student and teacher networks of SCRD but also increases the discrepancy of their abnormal features, achieving explicit constraint on the abnormal features of the student. Experimental results indicate that SCRD achieves outstanding anomaly localization performance, yielding more refined visualizations for anomaly localization.