Research on forest fire smoke object detection based on deep learning

Forest fire detection holds significant importance in protecting human life, property, and the ecological environment. In forest fire detection, smoke and fire detection are key identification criteria. Building on existing neural network-based models, this paper proposes a YOLO Firefly model derived from YOLOv8. This work optimizes YOLOv8 by incorporating the Firefly attention mechanism into its backbone, yielding the novel YOLO Firefly architecture to address the low detection rate of small flames. This optimization further enhances the model’s perception of multi-scale images. Experimental results show that YOLO Firefly attains 81.8% accuracy, outperforming YOLOv8n by 13.3% in detection performance. Notably, the model’s size is comparable to YOLOv8n, with a parameter count of 5.94MB. This combination of improved performance and compact size marks a significant leap in overall capability, making it better suited to forest fire detection tasks.