Few-Shot Learning Based on GAN and Model Fine-Tuning

Accurate identification of devices such as transformers, switches, transmission lines and poles is becoming increas <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{\gamma}$</tex> ingly important as the demand for automated monitoring and control in the power and energy industries increases. Traditional image classification methods often face performance bottlenecks due to insufficient samples and data imbalance. In this paper, we propose a few-shot learning based on data enhancement and model fine-tuning, which first generates images using conditional generative adversarial network (cGAN) to expand the training dataset, enhance the sample diversity, and solve the problem of uneven sample sizes in different categories. Subsequently, the pretrained ResNet50 model is fine-tuned according to the enhanced dataset to better adapt to the characteristics of power equipment. The experimental results show that after data enhancement and model fine-tuning, the classification accuracy of transformers and poles reaches 92% and 90% respectively, and the classification accuracy of switches and transmission lines is 88% and 85% respectively. The overall average accuracy is 89.75%, which proves the effectiveness of the model in the power equipment image classification task.