EdgeFireNet: a lightweight anchor-free fire detection network for edge devices

Early visual detection of fire and smoke plays a vital role in minimizing disaster impact, especially in resource-constrained environments such as UAVs, IoT sensors, and embedded surveillance systems. However, deploying accurate deep learning models on edge devices is challenging due to limitations in memory, compute, and power. We present EdgeFireNet, a lightweight, anchor-free object detector designed for single-frame fire and smoke recognition under edge computing constraints. With only 0.63 million parameters and 2.9 GFLOPs, EdgeFireNet achieves 65.4% mAP@0.5 on the D-Fire dataset, 85% of the accuracy of YOLOv5-Nano, while using just 25% of its parameters and 40% of the computation. EdgeFireNet incorporates several innovations for efficient detection: an Edge-Focus backbone with dynamic and edgeenhancing convolutions, a Graph Attention Fusion neck for lightweight cross-scale feature integration, and a dualfrequency IoU-aware head that improves class separation by exploiting frequency differences between fire and smoke. Confusion matrix analysis reveals strong classification performance with moderate confusion in background regions. Qualitative visualizations further demonstrate correct detections in typical scenarios and highlight limitations under visually ambiguous conditions. Overall, EdgeFireNet offers a compelling solution for real-time, low-power fire/smoke detection with high efficiency.