In response to the existing object detection algorithms are applied to complex fire scenarios with poor detection accuracy, slow speed and difficult deployment., this paper proposes a lightweight fire detection algorithm of Light-YOLOv5 that achieves a balance of speed and accuracy. First, the last layer of backbone network is replaced with SepViT Block to enhance the contact of backbone network to global information; second, a Light-BiFPN neck network is designed to lighten the model while improving the feature extraction; third, Global Attention Mechanism (GAM) is fused into the network to make the model more focused on global dimensional features; finally, we use the Mish activation function and SIoU loss to increase the convergence speed and improve the accuracy at the same time. The experimental results show that Light-YOLOv5 improves mAP by 3.3% compared to the original algorithm, reduces the number of parameters by 27.1%, decreases the computation by 19.1%, achieves FPS of 91.1. Even compared to the latest YOLOv7-tiny, the mAP of Light-YOLOv5 is 6.8% higher, which shows the effectiveness of the algorithm.