A Study of Attention-Based YOLOv5 for Forest Fire Detection

Fire detection is crucial for safeguarding lives and property. The development of deep learning technology has enabled significant potential in fire detection methods based on image and video analysis. The study proposes an improved fire detection algorithm based on YOLOv5 by incorporating a Coordinate Attention (CA) mechanism to enhance model performance. A dataset of 2,675 annotated flame images was constructed, with the CA module integrated into the C3 module of YOLOv5. Experimental results reveal that the enhanced model achieves a remarkable balance between precision and recall, with the F1 score peaking at 0.67774 near a confidence threshold of 0.13. The model maintains stable performance within the 0.1 to 0.5 confidence range, providing flexibility for practical applications. At higher confidence levels, the model exhibits nearly perfect precision. The integration of the CA mechanism with YOLOv5 significantly boosts fire detection accuracy and efficiency, offering new insights for developing more reliable fire safety systems.