FireFlow: A Multi-Scale Attention Convolutional Classification Network For Fire Monitoring

Fire monitoring is crucial for safeguarding life and property, traditional sensors and manual feature vision methods have problems such as narrow coverage, high false alarm rate, and underutilisation of multi-scale information, while existing convolutional neural network (CNN) models are difficult to be deployed in low-resource devices due to the large number of parameters and high computational cost. To this end, this paper proposes FireFlow, a multi-scale attention fire classification network, which consists of feature extractors and classifiers to achieve efficient capture of global context and local details of flames. The experiments are based on five publicly available datasets, including FireNet and SmokeFire, and the results show that FireFlow has a classification accuracy of $\mathbf{9 5. 7 6 \% - 9 9. 9 2 \%}$ on each dataset, which is significantly better than the comparative models, such as SqueezeNet and MobileNetV2, and provides a highly efficient solution for real-time fire monitoring of low-resource devices.