Few-shot image generation based on meta-learning and generative adversarial network

Abstract

Generative adversarial network (GAN) learns the latent distribution of samples through the adversarial training between discriminator and generator, then uses the learned probability distribution to generate realistic samples. Training a vanilla GAN requires a large number of samples and a significant amount of time. However, in practical applications, obtaining a large dataset and dedicating extensive time to model training can be very costly. Training a GAN with a small number of samples to generate high-quality images is a pressing research problem. Although this area has seen limited exploration, FAML (Fast Adaptive Meta-Learning) stands out as a notable approach. However, FAML has the following shortcomings: (1) The training time on complex datasets, such as VGGFaces and MiniImageNet, is excessively long. (2) It exhibits poor generalization performance and produces low-quality images across different datasets. (3) The generated samples lack diversity. To address the three shortcomings, we improved FAML in two key areas: model structure and loss function. The improved model effectively overcomes all three limitations of FAML. We conducted extensive experiments on four datasets to compare our model with the baseline FAML across seven evaluation metrics. The results demonstrate that our model is both more efficient and effective, particularly on the two complex datasets, VGGFaces and MiniImageNet. Our model outperforms FAML on six of the seven evaluation metrics, with only a slight underperformance on one metric. Our code is available at https://github.com/BTGWS/FSML-GAN.