Insulator Image Dataset Generation based on Generative Adversarial Network

For the intelligent processing of power equipment, its image datasets are often required as data support. However, the collection of power equipment image datasets is limited by the location and environment, and the number of collected datasets is relatively small, which cannot provide enough data for specific applications. This paper proposes to use image generation function of generative adversarial network to generate more images of electrical equipment from existing datasets with a smaller number, thereby increasing the size of electrical equipment datasets. In this paper, the insulator dataset is mainly used for experiment. First, the model building of the generative adversarial network is carried out. This paper uses pytorch to build three network model frameworks of generative adversarial networks, namely self-attention generative adversarial network, boundary equilibrium generative adversarial network and projected generative adversarial network, and selects relevant loss function and training method according to characteristics of each model. Second, models are trained on these three GAN s for the insulator image dataset and analyze experimental results of these three models. Finally, generated results and FID scores of the three models are compared. The FID score of projected generative adversarial network is the lowest, and the quality and diversity of the generated samples are the best, indicating that this model can better learn the characteristics of images and is more suitable for the generation of power equipment datasets.