Neural-symbolic framework for dynamic hazard detection through compositional visual reasoning

Abstract Industrial hazard detection faces significant challenges due to evolving safety regulations, diverse operational scenarios, and limited training data availability. Current approaches rely on predefined hazard categories and static safety rules, requiring extensive retraining when regulations change. To address these limitations, we propose a neural-symbolic framework that decouples hazard detection into two components: (1) a symbolic program representation of safety rules, and (2) a compositional visual reasoning engine based on Meta Module Networks. This separation enables dynamic adaptation to new safety standards without model retraining, while maintaining high detection accuracy. To evaluate our approach, we introduce HazardComp, a new dataset containing 2,006 real-world images annotated with scene graphs, safety rules, and corresponding hazard queries. The dataset spans diverse industrial environments and enables the evaluation across different hazard types. Experimental results demonstrate that our framework achieves higher performance compared to existing methods, with the inference time of 0.05 s and average accuracies of 86.9%, 90.4%, and 91.6% for object-based, relation-based, and logic-based hazard detection respectively. Our framework’s key innovation lies in its ability to reason about previously unseen hazard scenarios through compositional understanding, offering a more flexible and maintainable solution for real-world safety monitoring applications. Code and dataset are available at: https://github.com/hailingprojects/CVR-hazardDet/tree/patch-1 .