Aiming at the problem of low efficiency and accuracy of fireworks detection due to the small size of fire target and the confusion of fire feature with actual scene in complex environment, a small scale fireworks target detection method based on improved YOLOv5 is proposed. Firstly, a fourth detection layer is added to the third detection layer output in the original YOLOv5 model, so as to obtain a larger feature map for small target detection and strengthen the feature extraction capability of the network model. Secondly, in order to solve the problem that the target is prone to miss detection in the shielded scene, GIOU_Loss used to calculate the regression loss function of the target frame in the original network is replaced by DIOU_Loss. Finally, TensorRT is used to compress and accelerate the optimization of the model, and it is deployed to the Jetson TX2 development board for accelerated reasoning experiments. More fireworks scene data are constructed by replication enhancement method and a large number of experimental results show that the proposed method not only has a fast convergence speed, but also has a higher accuracy for small scale fire spot detection. It is suitable for popularization and application.