Reconstruction from a distilled encoder with edge-based pseudo-anomaly for industrial anomaly detection.

Industrial anomaly detection is an important research topic in the field of computer vision. Although widely studied, anomaly detection methods based on supervised learning have long faced challenges due to the scarcity of anomaly samples. To overcome this limitation, recent efforts have shifted toward reconstruction-based methods, which typically operate by first generating pseudo-anomalies and then reconstructing them. However, the pseudo-anomalies generated by current methods lack the requisite similarity and localization, and the reconstruction networks struggle to balance image fidelity with the accurate reconstruction of anomalous regions. To tackle these issues, this paper proposes an unsupervised anomaly detection framework called RDEAD. The core components of RDEAD are an Edge-based pseudo-anomaly generation strategy (EPA) and a distillation-based dual-encoder reconstruction network (YNet). EPA accurately generates pseudo-anomalies on the target object that closely resemble real anomalies in shape. YNet employs an encoder that has been distilled to provide the decoder with discrepancy features of anomalies, which are then used to reconstruct the anomalous regions accurately. Additionally, YNet incorporates an encoder feature fusion module (EFFM) to effectively integrate the features from dual encoders, enhancing detection performance. Experimental results on several widely used industrial datasets fully demonstrate the effectiveness of the proposed RDEAD method.