Towards efficient context-aware classification with compact VLM architectures: indoor fire case study

Accurate and reliable fire detection in indoor environments is critical for ensuring timely emergency responses and enhancing safety in both residential and industrial settings. While recent advances in deep learning have significantly improved fire and smoke detection, most existing systems remain limited to binary classification and often fail to distinguish between hazardous and benign fire events, leading to frequent and disruptive false alarms. To address this limitation, we propose a lightweight and efficient framework for context-aware fire classification. Specifically, input images are first encoded by a visual encoder and then interpreted by a Vision Language Model (VLM) to produce descriptive natural-language captions or semantic embeddings. A language model then processes these representations to perform high-level semantic classification into three categories: no fire, controlled fire (or fire under control), and dangerous fire. This design enables the system to reason about visual context and scene semantics, allowing for nuanced differentiation between visually similar but contextually distinct fire events. Our empirical results, evaluated on both re-labeled public datasets and our custom ConFire dataset, demonstrate that our approach achieves high accuracy while significantly reducing computational overhead. These findings highlight the effectiveness of integrating vision-language reasoning into fire classification tasks, paving the way for next-generation safety monitoring systems that are both intelligent and resource-efficient.