LFNet: Lightweight fire smoke detection for uncertain surveillance environment

Recently, many fire disasters took place all over the world. Therefore, the research communities pay more attention to the algorithms that can accurately detect fire disasters in the monitoring scene. Although the current detector algorithms achieve excellent performances in the benchmark object detection task, they can not achieve high recognition accuracy in fire monitoring tasks due to the fire disaster has no distinct rules in shape, color, and size. Moreover, the traditional learning-based object detection algorithms have limitations in model size and computational cost, so it can not be well applied in resource-constrained devices and Tactile Internet. In terms of the dataset, current learning-based object detection algorithms do not consider the impact of some fair weather. The models trained by such training sets are easily affected by the weather. We propose a new Lightweight model called LFNet, which has a small size and low computational cost to solve these problems. We also published a large-scale benchmark dataset to train the fire disaster detection model. To make the learning-based model trained by the proposed dataset possess higher robustness in various weather conditions, we utilize the atmospheric scattering model to simulate the original images into pictures affected by different degrees of simulated haze and dust for training. Experimental results show that the proposed LFNet can achieve state-of-the-art performance with limited model size. The presented data augmentation method can significantly improve the robustness of the current object detection model.