SA-YOLO: Stage-Aware Attention for Real-Time Fire and Smoke Detection on Edge Devices

초록

Fire incidents in modern industrial facilities and dense urban environments can spread rapidly, making detection within the initial golden time window critical for minimizing casualties and property damage. Conventional sensor-based fire detection systems are inherently limited by physical constraints, often resulting in delayed responses and exhibiting high miss rates in large-scale or high-ceiling environments. While vision-based detectors offer a promising alternative, lightweight models often lose representation power for small flames or low-contrast smoke under complex backgrounds. Although recent attention-centric detectors such as YOLOv12-n improve global feature modelling, their uniform attention design remains suboptimal for fire-smoke detection, where different visual cues dominate at different feature scales. This mismatch is particularly critical in fire-smoke detection, where flames and smoke exhibit fundamentally different scale-dependent visual characteristics that cannot be effectively handled by uniform attention refinement. To address this limitation, we propose Stage-Aware YOLO (SA-YOLO), a lightweight real-time detector that strategically allocates heterogeneous attention modules to backbone stages according to their representational roles. Specifically, Efficient Channel Attention (ECA) is deployed at the lower stage (P3) to preserve fine-grained details of small flames; Parallel Attention Module (PAM) is applied at the middle stage (P4) to enhance contextual cues around fire regions; and Residual Efficient Channel Attention (ResECA) is integrated at the higher stage (P5) to stably preserve weak and diffuse smoke signals that tend to vanish in deeper layers. Extensive experiments on the AI-Hub Fire Prediction Video dataset demonstrate that SA-YOLO consistently outperforms the YOLOv12-n baseline across single-class (fire-only and smoke-only) and multi-class (fire+smoke) settings. In the multi-class setting (100 epochs), SA-YOLO improves mAP@50 by 1.95 percentage points and Recall by 4.0 percentage points while preserving real-time inference speed. Despite the introduction of stage-specific attention modules, SA-YOLO maintains a throughput of 61.08 FPS with minimal computational overhead, confirming its suitability for edge-oriented fire monitoring systems operating under challenging visibility conditions.

키워드