In this paper, Investigated the characteristics of fire targets, the flames and smoke targets in natural light and nighttime ambient infrared thermal imaging environment, and incorporate an improved SE attention mechanism in the backbone feature extraction part of the YOLOv5s network to improve the network’s ability to detect and extract deep texture information and feature information; secondly, to address the problem of lack of integrity and occlusion in the case of fire spread, we choose Secondly, the YOLOv5s network model is optimized by using the CIoU loss function with better performance to address the problem of integrity and occlusion in the case of fire spread. The Ghost module is introduced to optimize the YOLOv5s + CIoU + SE-MA model in a lightweight way. And the recognition detection performance of the network models designed in this paper is compared on the self-built dataset. The results show that the detection FPS of The lightweight network model designed. In this paper achieves 87.7% mAP @ 0.5 reaches 37frames/s with 32% reduction in floating-point operations, which has excellent detection performance and system piggybacking capability in fire detection. Combined with infrared thermal imaging, the fires are realized for 24-h real-time monitoring function.