Answers are grounded in this paper and cite their sources.
Presents a novel anomaly detection framework for smart manufacturing, suggesting improved performance across domains.
Industrial anomaly detection plays a pivotal role in smart manufacturing; however, existing methods struggle with cross-domain data heterogeneity in diversified production lines. This study proposes Prototype-guided Two-branch Multi-class Anomaly Detection (PTMAD), a framework enabled by two synergistic innovations. First, we devise a multi-scale region reconstruction paradigm that replaces conventional image-level reconstruction. This is achieved by integrating a Vision Transformer-based hierarchical architecture with combinatorial masking strategies and visual token reconstruction objectives. Through this integration, our framework effectively captures multi-class features. In particular, the adaptive region masking mechanism preserves category-agnostic patterns and mitigates distribution shifts in classes. Second, an anomaly candidate screening pipeline is developed, comprising two phases: (1) a prototype memory-driven masking phase that generates high-confidence anomaly proposals by quantifying spatial deviations between test samples and class-specific prototypes; (2) a discriminative verification phase that analyzes local reconstruction fidelity through thresholded residual errors. This phase theoretically circumvents the ‘shortcut reconstruction’ dilemma, where anomalous regions are erroneously restored. Experiments on industrial benchmark datasets demonstrate the superiority of PTMAD.
No takes yet. Share an insight, caveat, or question.
Li et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: