Reproduce “H2VLR: Heterogeneous Hypergraph Vision-Language Reasoning for Few-Shot Anomaly Detection” on OpenResearch

About this paper

Jianghong Huang, Luping Ji, Weiwei Duan, Mao Ye

H₂VLR introduces a heterogeneous hypergraph vision-language reasoning framework for few-shot anomaly detection, reformulating the task as a high-order relational inference problem. The framework achieves leading average AUROC scores for image-level classification (90.37% for 4-shot) and pixel-level localization (95.68% for 4-shot) on industrial datasets, outperforming existing VLM-based methods and exhibiting competitive performance with full-sample approaches.

Key insights

• Reformulating few-shot anomaly detection as a high-order relational inference problem using a heterogeneous hypergraph explicitly captures complex structural dependencies and global contextual consistency beyond traditional pairwise matching.
• Dynamically inducing task-adaptive textual semantics from limited support images provides more specific and robust VLM priors than generic prompts, enhancing anomaly discrimination.
• Combining a coarse base anomaly map (from global semantic priors) with a refined structure-aware residual map (from hypergraph reasoning) through a fusion mechanism significantly improves both detection robustness and localization precision.

Summary by alphaXiv.