A Faster Fire Detection Network with Global Information Awareness

Jinrong Cui, Haosen Sun, Min Zhao, Ciwei Kuang, Yong Xu

Abstract: A fast fire detection can help prevent the further loss of life property. Existing fire detection methods often concentrate into two directions. Some focus on building models with Transformer to perceive the global information of fire for higher accuracy, while others working on optimizing the model’s size to make it more lightweight. However, all these methods suffer from a certain loss in detection speed. Therefore, in this paper, we present a faster fire detection network with global information awareness (FasterGA-Net). Specifically, to enable the fire detection network to have awareness of fire’s global information, the UniRepLKNet Block based on large kernel convolution is adopted into our model. With a lower computational complexity than Transformer based module, this module avoids severe drop in detection speed. Besides, a lightweight convolution operator PSConv is designed to build the efficient feature fusion network in the neck, further improving our network’s detection speed. Extensive experiment results show that, our proposed model achieves the highest accuracy among all comparative models while having a faster detection speed than the baseline model.