Feature Consistency Learning for Anomaly Detection

Anomaly detection in industrial images is a challenging automated vision inspection task. Given an input image, it is essential to know not only whether or not it is abnormal, but also to locate the anomaly. Currently, knowledge distillation-based teacher-student (T-S) networks have nearly saturated image-level anomaly detection on publicly available datasets, however, pixel-level anomaly localization is still very tough. To address this problem, we propose a spatial neighboring coding (SNC) module that facilitates the localization of anomalies by encoding the contextual information of the neighborhood of each element in feature space. Our SNC can be easily plugged into diverse T-S network-based anomaly detectors to improve their performance. Subsequently, we propose a feature consistency learning (FCL) method that learns low-level and high-level feature consistencies in a unified framework for anomaly detection tasks. The proposed FCL is capable of achieving more accurate anomaly localization by imposing consistency constraints on the features extracted from the T-S network. Experimental results on benchmark datasets for industrial image anomaly detection show that our FCL method achieves image-level Area Under Receiver Operator Characteristic Curve (AUROC) of 99.1% and 98.4%, pixel-level AUROC of 97.7% and 99.0%, and pixel-level Area Under Per-Region Overlap (AUPRO) of 97.4% and 95.5% on MVTec AD and VisA datasets, respectively, demonstrating the effectiveness of our FCL.