Normality-Anchored Regional Evidence Reasoning for Zero-Shot Industrial Anomaly Detection

Abstract

Industrial visual inspection is a key task in intelligent manufacturing and quality control. However, defective samples in real production lines are usually scarce, diverse in appearance, and expensive to annotate, which makes supervised models that rely on large numbers of defective samples difficult to adapt to new products and unknown defect scenarios. To address this problem, this paper proposes NARE-ZSAD, a normality-anchored regional evidence reasoning framework for zero-shot industrial anomaly detection. The proposed framework first decomposes an input image into local candidate regions through a language-guided multi-scale region proposal mechanism, and then performs region-level vision-text alignment using normal and abnormal textual prompts, thereby reducing the interference caused by global image semantics and large normal background areas. Subsequently, a normality memory bank is constructed to characterize local normal visual patterns of the target product, reformulating anomaly detection as a joint estimation of semantic anomaly responses and deviations from normal patterns. Furthermore, a text-guided pseudo-defect evidence calibration strategy is designed to synthesize local anomaly evidence, such as scratches, stains, missing areas, and color shifts, on defect-free images. This strategy calibrates abnormal prompt responses and regional thresholds without using real defective samples. Finally, a conflict-aware expert reasoning module integrates semantic anomalies, normality deviations, texture inconsistencies, and prompt uncertainty to output anomaly scores, anomaly localization maps, triggered prompts, and manual review decisions. Unlike conventional GAN-based defect image augmentation methods, NARE-ZSAD does not generate full defective images to train supervised classifiers. Instead, it enables region-level evidence discovery, normality-deviation reasoning, and reliability-aware decision making under zero-shot conditions, providing an interpretable and reviewable solution for industrial inspection in defect-scarce and unknown-anomaly scenarios.