Feature Retention Guidance-Based Asymmetric Distillation Network for Industrial Precision Surface Defect Detection

Industrial precision surface defect detection is extremely challenging and pivotal in industrial manufacturing processes because of unforeseen and diverse anomalies. Recently, unsupervised anomaly detection methods based on knowledge distillation and feature similarity have been developed and have shown remarkable potential. Although most current approaches focus on knowledge distillation and feature similarity, they prove insufficient in detailed and precise detection tasks where omission and commission errors frequently occur. To address this issue, this study proposes a novel feature retention guidance-based asymmetric distillation network (FRADN) paradigm that effectively improves its discriminative capacity during the inference process by fully considering knowledge distillation and feature similarity. First, a novel knowledge distillation-based module considering relational distillation named the local feature retrieval module (LFRM) is introduced to precisely detect local defects. Second, a novel knowledge distillation-based module with global feature fusion named the global feature retrieval module (GFRM) is introduced to detect global defects from a broad contextual viewpoint. Furthermore, a novel feature retention guidance module (FRGM) is introduced to focus on critical information and refine the perception capacity of the network. On the MVTec AD dataset, the FRADN achieves a state-of-the-art pixel-level area under the per-region overlap curve (AUPRO) of 96.6% and an image-level area under the receiver operating characteristic curve (AUROC) of 99.4%. Extensive experiments conducted on mainstream anomaly detection datasets show that the FRADN outperforms the state-of-the-art competitors in terms of accuracy and the experiments on real-world industrial PCBA product datasets prove the practical applicability of our method.