Generative adversarial network based on frequency domain data enhancement: Dual-way discriminator structure Copes with limited data

Abstract

The excellent image-generation ability of generative adversarial networks (GANs) has been widely used. However, training a GAN requires large-scale data support, which hinders in-depth development. Therefore, the research on stable training of GANs under limited data conditions is helpful to further expand the application scenarios. To solve this problem, a new network based on a dual-ways discriminator structure is designed, used to eliminate the problem that a single discriminator model is prone to overfitting under the condition of limited data. Then, the problem that the traditional data augmentation strategy is limited to pixel space and lacks attention to the overall structure and contour of the image is analyzed. An adaptive dynamic data augmentation strategy based on the Laplace convolution kernel is proposed from the frequency domain space, which realizes the purpose of implicitly increasing the training data in the training process. This new designed module improves the performance of the generative adversarial network. Through extensive experiments, it was confirmed that the new network, named FD-GAN, achieved prefer image generation ability, and its Fid score reached 4.58, 12.007, and 10.382 in the AFHQ-Cat, AFHQ-Dog, and TankDataSet datasets, respectively.