A novel approach based on convolutional neural networks ensemble for fire detection

Fire is a severe catastrophe that harms consequences for humans, the environment, materials, and the economy. A practical solution to fighting fires involves using early warning systems that enable timely detection, thereby reducing their spread and mitigating possible risks. With technological development, surveillance cameras are available in most places, encouraging researchers to develop fire detection systems using convolutional neural networks (CNN). However, most researchers use very deep CNNs, which are computationally expensive, requiring large memory resources and robust hardware for training. To bridge these gaps, four new simple and lightweight CNN networks to classify fires and detect them in their initial stage are proposed in this paper, CNN_A, CNN_B, CNN_C and CNN_D. Furthermore, to improve the algorithm's performance and reduce the error rate, a new assembly approach that combines the predictions of the four networks is applied to single models. A comparison of the proposed algorithms with the recent works on fires shows the efficiency of the approach. Above all, the CNN_BC ensemble model that reached an excellent accuracy of 99.69% and 96.01% on two challenging testsets, achieving a reduced number of floating-point operations and a smaller size. Therefore, the CNN_BC model has an excellent robustness and can easily classify fire at the early stages.