Industrial Zero-Shot Anomaly Detection (ZSAD) aims to identify defects in industrial images without requiring training samples from target categories, making it highly suitable for real-world applications such as product quality inspection in manufacturing. However, existing methods predominantly rely on manually designed hard-text prompt templates, which struggle to adapt to diverse anomaly features and exhibit limited generalization capabilities. To address the problem, a hierarchical hybrid softened prompt learning for instance-aware ZSAD (H 2 SP-AD) is proposed, in which the prompts are softened and visually individualized. Specifically, a hard prompt softening strategy is designed to insert the learnable soft-prompt vectors into generic hard-prompt templates to capture common anomaly patterns across categories. Furthermore, instance-aware learnable parameters are introduced to adaptively adjust softened prompts based on individual features, mitigating overfitting to auxiliary data. Simultaneously, a learnable local tokens are integrated into the visual encoder to enhance anomaly region perception. Experiments on seven widely-used industrial datasets demonstrate that H2SP-AD achieves state-of-the-art performance in both anomaly classification and localization tasks, achieving 14.3% improvement in pixel-wise AUROC over hard prompt (WinCLIP) and 2.1% in image-wise AUROC over soft one (AnomalyCLIP), showing strong potential for practical deployment in smart manufacturing systems.
Yang et al. (Thu,) studied this question.