View Fusion Based Zero-shot Multi-view Anomaly Detection

초록

This paper presents a View Fusion-based Zeroshot Multi-view Anomaly Detection framework that leverages pretrained vision-language models such as CLIP. To address the challenge that object defects may be visible in some views but occluded in others, we propose a Common-view Feature Block (CVF-Block) that extracts shared representations from multiview images through a VAE-based bottleneck without additional training. The extracted common-view features are integrated into CLIP's visual encoder via a residual connection, enabling better reasoning across multiple viewpoints. Experiments on the RealIAD dataset demonstrate that the proposed method outperforms existing single-view zero-shot approaches, achieving improved robustness and detection accuracy in multi-view anomaly detection. © 2026 IEEE.

키워드