跳转至

文章背景与核心概要

本文探讨了针对安全关键型火灾检测系统，如何将领域专用的视觉语言模型（VLM）进行压缩，以实现完全的端侧（On-Device）部署。尽管视觉语言模型能够通过对场景上下文的推理来提高准确性并减少误报，但其庞大的模型体积使其无法部署在标准的嵌入式传感器上。

为了解决这一挑战，作者提出了一种教师-学生知识 distillation（知识蒸馏）框架，将针对火灾理解微调过的大型 VLM 压缩为轻量级学生模型。跨多个 VLM 家族和模型规模的实验表明，精简后的学生模型保留了其教师模型的大部分火灾理解能力。当部署在商用 Detectium 传感器上时，蒸馏后的模型（尤其是 Qwen2.5-0.5B）在推理准确性、延迟、内存占用以及失效模式管理之间取得了最佳的平衡，为在资源受限和安全关键环境部署领域专用 VLM 提供了更广泛的指导。

Distilling Vision-Language Models for On-Device Fire Understanding

Distilling Vision-Language Models for On-Device Fire Understanding

Summary

Summary

This paper explores the compression of domain-specialized vision-language models (VLMs) for fully on-device deployment in safety-critical fire detection systems. While VLMs improve accuracy by reasoning about scene context and reducing false alarms, their large size prevents deployment on standard embedded sensors. Using a teacher-student knowledge distillation framework, the authors compress large fire-understanding VLMs into lightweight student models. Experiments demonstrate that compact students retain most of their teachers' capabilities. When deployed on commercial Detectium sensors, the distilled models—particularly Qwen2.5-0.5B—achieve an optimal balance between reasoning accuracy, inference latency, memory usage, and failure mode management for resource-constrained environments.

Document Metadata

Document Metadata

Metadata Field Details
arXiv Identifier arXiv:2609.05782 [cs.AI]
Subjects Artificial Intelligence (cs.AI)
Authors Mohammad Kazzazi, Zixuan Liu, Siavash Khajavi
Submission Date September 5, 2026
License Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International license icon
Metadata Field Details
arXiv Identifier arXiv:2609.05782 [cs.AI]
Subjects Artificial Intelligence (cs.AI)
Authors Mohammad Kazzazi, Zixuan Liu, Siavash Khajavi
Submission Date September 5, 2026
License Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International license icon

Abstract

Abstract

Vision-language models (VLMs) offer a promising alternative to conventional fire detection systems by reasoning about the semantic context of a scene and thus reducing false alarms, yet their large model size makes deployment on embedded fire sensors impractical. In this paper, we study how domain-specialized VLMs can be compressed for fully on-device deployment without losing the safety-critical behavior required for fire detection. We develop a teacher-student knowledge distillation framework in which large VLMs fine-tuned for fire understanding can be distilled into lightweight students. Experiments across multiple VLM families and model scales show that compact students preserve most of their teachers' fire-understanding capability. We further deploy the distilled models on our commercial Detectium fire detection sensor and jointly evaluate reasoning accuracy, latency, and memory usage. The results show that compression and deployment affect not only accuracy but also model failure modes, with Qwen2.5-0.5B providing the strongest overall deployment trade-off. Our findings provide broader guidance for deploying domain-specialized VLMs in resource-constrained, safety-critical settings.

Full-Text & Access Links

Full-Text & Access Links

References & External Tools

References & External Tools