Runming Wang

H​W​∑i,j(|Si+1,j−Si,j|+|Si,j+1−Si,j|).\mathrm{TV}(S)=\frac{1}{HW}\sum_{i,j}\big(|S_{i+1,j}-S_{i,j}|+|S_{i,j+1}-S_{i,j}|\big). (9) We set r=0.01r=0.01 by default. Train-good statistical calibration. We fit normal statistics of ϕ​(S)\phi(S) using train-good samples only. Let μ∈ℝ3\mu\in\mathbb{R}^{3} and σ∈ℝ3\sigma\in\mathbb{R}^{3} denote the per-dimension mean and standard deviation of ϕ​(S)\phi(S) over train-good anomaly maps. At test time, we compute a diagonal Mahalanobis (standardized Euclidean) distance: Dstruct​(S)=‖ϕ​(S)−μσ+ε‖2.D_{\mathrm{struct}}(S)=\left\lVert\frac{\phi(S)-\mu}{\sigma+\varepsilon}\right\rVert_{2}. (10) This structural score depends only on the anomaly map SS and does not modify patch scores or localization. Automatic scale matching (auto-λ\lambda). Because SbaseS_{\mathrm{base}} and DstructD_{\mathrm{struct}} can have different scales, we compute a category-wise automatic weight using train-good statistics: λauto=Std​(Sbase)Std​(Dstruct)+ε,\lambda_{\mathrm{auto}}=\frac{\mathrm{Std}(S_{\mathrm{base}})}{\mathrm{Std}(D_{\mathrm{struct}})+\varepsilon}, (11) where the standard deviations are computed over train-good samples. The final hybrid image-level score is Shyb​(x)=Sbase​(x)+λauto⋅Dstruct​(S​(x)).S_{\mathrm{hyb}}(x)=S_{\mathrm{base}}(x)+\lambda_{\mathrm{auto}}\cdot D_{\mathrm{struct}}(S(x)). (12) For analysis, we also report a struct-only score Sstruct​(x)=Dstruct​(S​(x))S_{\mathrm{struct}}(x)=D_{\mathrm{struct}}(S(x)). StructCore modifies only the image-level score and preserves pixel-level localization behavior. 4 Experiments 4.1 Experimental Settings Benchmarks and Protocols. We evaluate StructCore on two standard industrial UAD benchmarks, MVTec AD (15 categories) and VisA (12 categories). Following the one-class protocol, all components are fitted using only normal (train-good) samples, and evaluation is performed on test sets containing both normal and anomalous images. Evaluation Metrics. We report both detection and localization metrics. For image-level detection, we report I-AUROC and I-AP, and additionally report I-F1-max (the best F1 over thresholds) as a reference. For pixel-level localization, we report P-AUROC, P-AP, and P-F1-max when pixel annotations are available, and additionally report AUPRO on MVTec AD. Implementation and Backbone Details. Feature extraction: We use a frozen DINOv2 ViT-B/14 backbone (patch size 14) [32]. Preprocessing: We follow the preprocessing protocol of Dinomaly [17]: images are resized to