VGCAD: Visual Geometry‐Guided Circulant Attention Distillation for Industrial Anomaly Detection

This is the authors' abstract. We don't add key points for this paper.

Industrial visual anomaly detection aims to identify and localize unknown defects using only normal samples for training. Efficient teacher–student methods reduce inference cost by avoiding large‐scale feature‐memory retrieval, but their compact reconstruction paths often provide limited modeling of spatial relations among visual tokens. This limitation is critical for industrial inspection, where defects commonly appear as subtle disruptions of regular textures, repeated layouts, or local geometric structures. To address this problem, this paper proposes Visual Geometry‐Guided Circulant Attention Distillation (VGCAD), an efficient visual anomaly detection model. VGCAD introduces a sine‐position‐enhanced circulant attention module to strengthen global token‐correlation modeling with explicit two‐dimensional spatial information, enabling the student decoder to reconstruct teacher features from structure‐aware token representations. In addition, VGCAD employs a high‐level geo‐semantic manifold consistency constraint to overcome the limitation of pointwise cosine or mean squared error (MSE) distances in measuring abstract semantic features. Experiments on MVTec AD, VisA, and MPDD show that VGCAD improves image‐level anomaly detection and pixel‐level localization while maintaining fast inference comparable to efficient baseline models. Extensive ablation and visualization results further verify the effectiveness of the proposed circulant attention module and geo‐semantic manifold constraint. The source code of the proposed method is publicly available at https://github.com/CreatedTRYNA/VGCAD .