Abstract
Reconstruction-based methods leveraging Vision Transformers (ViT) have demonstrated significant potential in unsupervised industrial anomaly detection. However, existing single-stream architectures struggle with an inherent trade-off when processing multi-scale features: while high-resolution features enhance pixel-level localization, they often introduce local noise that degrades image-level detection performance. To resolve this conflict between global semantic stability and local detail sensitivity, we propose a dual-encoder framework named GLM-AD (Global-to-Local Modulation for Anomaly Detection). In which, the global encoder captures stable structural semantics, while the local encoder focuses on fine-grained defect-sensitive patterns. On top of this architecture, we introduce a Global-to-Local Feature Modulation (GLFM) module, which performs channel-wise modulation to implicitly constrain local feature reconstruction using global semantic cues. Furthermore, to cope with the significant morphological heterogeneity of industrial anomalies, we propose a Multi-Scale Ensemble Scoring (MSE-Scoring) strategy. As a model-agnostic and label-free post-processing scheme, MSE-Scoring integrates anomaly responses across multiple observation profiles, effectively alleviating the signal-to-noise imbalance inherent in high-resolution features. Extensive experiments conducted on three public industrial benchmarks(MVTec-AD, VisA, and Real-IAD) demonstrate that GLM-AD consistently achieves state-of-the-art performance in terms of detection robustness and localization accuracy, validating its effectiveness and practical applicability in real-world industrial inspection scenarios.

This publication has 10 references indexed in Scilit: