WFDRNet: Wavelet-Embedded Feature Distillation and Refinement Network for Unsupervised Anomaly Segmentation

We propose a novel Wavelet-Embedded Feature Distillation and Refinement Network (WFDRNet) to address the limitations of existing unsupervised anomaly detection methods in local anomaly localization. WFDRNet combines the advantages of Reverse Knowledge Distillation (RKD) and Autoencoder (AE) while introducing Discrete Wavelet Transform (DWT) to enhance the model’s sensitivity to anomalous regions, thus improving localization accuracy. The network employs a two-stage training strategy: in the first stage, the student network learns feature representations from the teacher network; in the second stage, the Anomaly Mask Generation Module (AMG) uses multi-scale features for anomaly localization. Experimental evaluations on the MVTec AD dataset demonstrate that WFDRNet surpasses existing state-of-the-art (SOTA) methods in Image-AUROC, Pixel-AUROC and Pixel-AP, particularly in precise anomaly localization, validating its effectiveness in complex contexts. Code is available at: https://github.com/IvanZhu666/WFDRNet.