Aiming at the problem that the power line foreign object recognition model can use fewer data sets, and the traditional SinGAN model generated data does not match the foreign object recognition model, the quality is poor, and it takes too long, the improved SinGAN model is proposed. Based on the improved SinGAN model, an affine transformation unit and a size transformation unit are added to further enhance the data set, and an image filtering unit is added to improve the data quality required by the power line foreign object recognition model. By improving the SinGAN back propagation training process and SinGAN’s singleprecision generator structure, the quality of model generation is improved and the time spent is reduced. Experimental results show that after 50 experiments, the average Frechet starting distance score (FID) of the improved SinGAN is 91375, and the average training time is 121 h. Compared with traditional SinGAN, it is reduced by 27247% and 8731% respectively. Compared with other mainstream generative adversarial networks, improved SinGAN has better foreign object data generation capability,which can enhance the data required by the power line foreign object recognition model, and has superiority.