Forest fire smoke recognition and detection based on EfficientNet

Forest fire, which is particularly difficult to put out, will result in lots of loss. In addition, it will also have a negative impact on the economy and the environment. Thus, the research on forest fire and smoke detection is always able to attract people’s attention. This paper combines forest fire recognition with deep learning. As a light neural network, EfficientNet is able to avoid the complexity and blindness of manual feature extraction in traditional image recognition methods. In this paper, Whale Optimization Algorithm, NAS structure search, and Progressive Learning Strategy are used to optimize the EfficientNet. Firstly, the optimal scaling factors was found. Secondly, the EfficientNet+ was designed by imitating the structure search strategy of MnasNet, and using NAS to search for the most spatial structure. Finally, the model was trained with a progressive learning strategy. The optimized EfficientNet+ has a faster training speed, higher accuracy, and fewer parameters. In terms of inference speed, it has higher accuracy and faster speed. Compared to the EfficientNet, the training time is 10h, the inference speed is 25ms, and the model accuracy is improved by 1% after optimization (85.3% vs. 86.8%).