YOLOv5-based Flame Detection Algorithm Research

Fire disasters are common and pose significant threats to the safety of people’s lives and property, making fire detection and early warning systems crucial. Traditional sensorbased fire detection methods suffer from slow response times, while image-based fire detection techniques can effectively address this issue. In this paper, we focus on the challenges posed by the fuzzy boundaries and irregular shapes of flames. We introduce the SIoU loss function into the YOLOv 5 model and incorporate the ECA attention mechanism into its Backbone to enhance the model’s ability to extract flame region features and reduce localization errors. The improved model outperforms the original YOLOv5, achieving a $12.8 \%$ increase in recognition accuracy on the test set. This demonstrates that the enhanced model offers higher precision and holds promising potential for practical applications.