Zero-shot industrial defect detection via domain-adaptive physical priors and semantics-guided Bayesian inference

Abstract

Automated visual inspection serves as a cornerstone of intelligent manufacturing, playing a critical role in ensuring product quality and production efficiency. However, general Vision-Language Models (VLMs) often suffer from significant performance degradation when applied to industrial zero-shot anomaly detection due to severe domain shifts and the deficiency in fine-grained semantic adaptation. To address these challenges, this paper proposes a cascaded detection framework fusing physical priors with semantic reasoning. The framework first incorporates a physically constrained geometric domain adaptation module, employing YOLO-based alignment and M-estimation-based robust background suppression to eliminate non-semantic interference. Subsequently, an inference-time explicit semantic injection strategy is introduced to deterministically map domain expert knowledge into conditional state vectors within the Bayesian latent space. Furthermore, a semantics-guided Bayesian inference mechanism is constructed to jointly model the posterior distribution of visual contexts and injected semantics via residual cross-attention, achieving precise pixel-level localization. Experimental results on the self-constructed BottleCap-Seg dataset demonstrate that the proposed method achieves an image-level AUROC of 97.48% and a pixel-level Average Precision (Pixel-AP) of 37.25%, significantly outperforming current representative methods and validating the effectiveness of the proposed vertical domain adaptation strategy.