Data Generation Using Generative Adversarial Network with Twin Normalization

Abstract Traditional generative adversarial networks require a large number of samples during training. When generating high-resolution images, it may suffer to mode collapse, resulting in lower quality outputs. This paper proposes a generative adversarial network data generation method (Twin-Norm-GAN) based on twin normalization, which is expected to relief the problem of mode collapse when generating high-resolution images with few samples. The method first normalizes the gradient of the discriminator, that is, imposes gradient norm constraints, which further prevents the phenomenon of mode collapse; then adds a normalized attention module to the generator and discriminator to suppress less significant weights, a weight sparsity penalty is applied to the attention module, which improves the quality of the generated images. Experiments are conducted using standard datasets such as FFHQ, Panda, and art paintings, and compared with the current state-of-the-art GAN model (FastGAN) for generating high-fidelity images with few samples through qualitative and IS, FID quantitative evaluation methods, the experiments show that compared with the contrast model, this method can generate higher quality images. And the classification experiments prove that the data generated by this method are available for the training of downstream task models.