Heterogeneous Multi-Score Integration for Structural and Logical Anomaly Detection

Unsupervised anomaly detection identifying both structural and logical anomalies is essential for modern industrial manufacturing. Although multi-score frameworks have been proposed, existing methods tend to reuse identical intermediate outputs across multiple branches. This structural redundancy limits the potential benefits of mutual complementarity. To mitigate this issue, this study proposes a novel multi-score framework integrating three branches operating on distinct mechanisms. These include a reconstruction branch based on the Adaptive Mask-Inpainting Network (AMI-Net), a segmentation branch utilizing Patch Histograms, and a feature branch employing Local-Global Student-Teacher (LGST) interactions. This work presents the following two technical contributions. First, the anomaly scores acquired from the heterogeneous branches are aligned through statistical normalization derived from the normal validation dataset before weighted fusion. Second, a Structural Similarity (SSIM) loss is introduced to the inpainting network of AMI-Net to enhance structural anomaly detection. Experiments on the MVTec Logical Constraints (LOCO) benchmark demonstrate that the proposed framework achieves an image-level Area Under the Receiver Operating Characteristic (AUROC) of 96.5%, with 98.4% for logical anomalies and 94.5% for structural anomalies, outperforming other solutions.