HRD-Net: A Hybrid Residual Denoising Network for Anomaly Detection and Localization Based on Knowledge Distillation
Jinke Liu, Jian Wang, Yihang Gao, Ken Deng, Fang Lv · 2025
Knowledge distillation has exerted a considerable influence in the domain of anomaly detection. In the traditional knowledge distillation architecture, due to the excessive similarity between the student network and the teacher network, the problem of overfitting is prone to occur. Hence, this paper presents a knowledge distillation anomaly detection and localization network based on hybrid dilated residual denoising (HRD-Net). Firstly, a hybrid dilated residual denoising module (HDRDM) is proposed, employing dilated convolution (s-DConv) and the channel attention mechanism to enhance the feature denoising capability of the student network, thereby preserving the deep semantic information of normal samples while suppressing the interference of abnormal features. The anomaly synthesis module (ASM) synthesizes anomalous samples through randomly generated masks and Poisson fusion technology to alleviate the scarcity of abnormal samples in practical applications and enhance the generalization ability of the model. The experimental results on the MVTecAD dataset demonstrate that HRD-Net has achieved excellent detection and localization performance in image-level anomaly detection and pixel-level anomaly localization, with the performance reaching 97.3%, 97.5%, and 93.1% respectively in terms of average image-level AUC, pixel-level AUC, and PRO.