Multi-scale Forest Flame Detection Based on Improved and Optimized YOLOv5

Forest fire may have very serious impacts on the natural environments and human beings. It is important to detect the source of a fire before it spreads. The existing flame detection algorithm has the problem of weak generalization and not fully considering the influence of flame target size on detection. To enhance the ability of flame detection of different sizes, ground flame data and UAV forest flame data were combined in this study. Cosine annealing algorithm, label smoothing and multi-scale training were introduced in order to improve the detection accuracy of the model. The experimental results show that our improved YOLOv5 has strong generalization and good detection effect for different sizes of flames. The mAP50 value of the improved YOLOv5 reaches 93.8%, which is 7.4% higher than YOLOv5 (mAP50) and 14.8% higher than YOLOv5 (mAP95). The proposed model has the advantages of strong generalization and low false detection rate, and has high detection accuracy for flame targets at different scales.