Aiming at the problems of the traditional sensor′s late detection of fireworks and its inability to give details of fireworks, as well as the imbalance between detection efficiency and accuracy of the current mainstream fireworks detection algorithms, an improved YOLOv5s light detection algorithm for fireworks was proposed. The second convolutional module in Backbone is replaced with Stem module, which can improve the model′s detection performance of small target space information and effectively control the total floating point operand. C3Ghost module and Ghost convolution module are introduced in Backbone and Neck to reduce the number of network parameters and improve the performance of fireworks detection. In order to distinguish the importance of different features in the process of feature fusion, a structure of adding learnable weight parameters to PAN is proposed, which significantly improves the average accuracy of fireworks detection. The experimental results show that compared with the original model, the weight of the model is reduced from 14.4 M to 10.2 M, GFLOPs is reduced from 15.8 to 3.7, and the average accuracy is increased by 1.1%. The improved model has improved the performance of pyrotechnic detection while being lightweight.