Abnormal Discrepancy-Guided Knowledge Distillation for Image Anomaly Detection

Knowledge distillation is a cornerstone of image anomaly detection for amplifying subtle defects via teacher–student discrepancy, yet existing methods rely on feature alignment loss that causes reconstruction error confusion and degrades accuracy. To address this critical limitation, this study proposes an abnormal discrepancy-guided knowledge distillation method (DiffKD) that differentially guides student feature reconstruction through channel-level discrepancy masks, leveraging normal features as supervisory signals and abnormal discrepancy features as constraints to enhance anomaly detection performance. The approach integrates a knowledge distillation network for feature reconstruction with a segmentation network for anomaly localization, while utilizing prior anomaly samples and synthetic anomaly samples to provide real-time training data of anomalous samples. Extensive evaluations on the SUT-Crack and MVTec AD benchmarks validate the effectiveness and generalizability of our approach. On MVTec AD, it achieves 80.7% average precision (AP) and 81.9% instance-level average precision (IAP), showing competitive performance against the representative methods evaluated under the same protocol. These results not only demonstrate significant improvements in IAD accuracy but also highlight its promise for enabling real-time, automated anomaly detection in practical applications.