Using Multimodal Large Language Models for False Alarm Reduction in Image-Based Fire Detection
This study proposes Flash-Cascade, the first sub-second MLLM-based firewall that leverages CoT for efficiently filtering false alarms, and fine-tuned Qwen2-VL-7B-Instruct on a multi-granularity instruction dataset via Low-Rank Adaptation to enable robust performance even without explicit CoT guidance.