KD-ADSeg: dual-stage knowledge distillation with segmentation-guided asymmetric denoising for industrial anomaly detection

Anomaly detection in industrial production is critical for ensuring product quality and operational safety. However, anomalous images often exhibit multi-scale structures, fine-grained details, and blurred boundaries, which challenge existing methods in localization accuracy and boundary delineation. To address these issues, this paper proposes a Dual-Stage Knowledge Distillation with Segmentation-Guided Asymmetric Denoising framework for Industrial Anomaly Detection (KD-ADSeg), aiming to enhance discriminative ability for anomalous regions and improve separation and localization under complex backgrounds. Specifically, an Asymmetric Denoising Student Network (ADSNet) is introduced and trained on synthetically generated anomalous images to suppress background noise. In addition, a Residual Local Importance-Aware Block (ResLIA-Block) is designed to integrate local attention with multi-scale residual aggregation, improving feature representation for segmentation. Furthermore, a Context-Aware Multi-path Propagation (CAMP) module is proposed to enhance contextual modeling and improve detection of blurred boundaries and large-scale defects through multi-scale feature fusion. Extensive experiments on the Real-IAD dataset show that KD-ADSeg achieves significant improvements in both pixel-level and image-level anomaly detection, demonstrating its effectiveness in industrial inspection. Its generalization ability is further validated on the MVTec AD, 3CAD, and VisA datasets.