Image Anomaly Detection via Knowledge Distillation With Cross‐Anomaly Sample Augmentation

Current image anomaly detection techniques mainly focus on using normal samples throughout the training phase to minimize the model's reliance on anomaly samples, overlooking the valuable information contained in anomalous samples that are often available in real‐world scenarios. However, lacking guidance from real anomaly samples, the model learns a decision boundary based exclusively on normal data, which inevitably results in an ambiguous boundary and poor generalization, rendering the model vulnerable to complex background interference and prone to high false alarm rates. To address this limitation, we propose a Few‐Shot Prior anomaly driven (FSPD) method that leverages a blend of prior anomalous samples and synthetic anomalies based on normal samples to enhance model training. Our approach integrates a knowledge distillation network for accurate feature reconstruction and a segmentation network for precise anomaly localization. Comprehensive experiments on the SUT‐Crack and MVTec AD datasets showed that our method achieved superior segmentation performance, with an AP of 81.7% and an IAP of 81.6% on MVTec AD, surpassing state‐of‐the‐art methods. By effectively utilizing both normal and anomalous samples, our approach ensures more accurate and efficient anomaly detection.