MRD-GAN: Multi-representation discrimination GAN for enhancing the diversity of the generated data

Abstract

The generative adversarial network (GAN) is a highly effective member of the generative models category and is extensively employed for generating realistic samples across various domains. The fundamental concept behind GAN involves two networks, a generator and a discriminator, competing against each other. During the training process, generator and discriminator networks encounter several issues that can potentially affect the quality and diversity of the generated samples. One such critical issue is mode collapse, where the generator fails to create varied samples. To tackle this issue, this article introduces a GAN approach called the multi-representation discrimination GAN (MRD-GAN). In this approach, the discriminator supports concurrent network discrimination flows to manage different representations of the data through various transformation functions, such as dimension rescaling, brightness adjustment, and gamma correction applied to the input data of the discriminator. We use a fusion function to aggregate the output of all flows and return a consolidated loss value to update the generator's weights. Hence, the discriminator conveys diverse feedback to the generator. The proposed approach has been evaluated on four distinct benchmarks, namely CelebA, Cifar-10, Fashion-Mnist, and Mnist. The experimental results demonstrate that the proposed approach surpasses the existing state-of-the-art GAN models in terms of FID metric that measures the diversity of the generated samples. Significantly, the proposed approach demonstrates remarkable FID scores of 14.02, 30.19, 9.42, and 3.14 on the CelebA, Cifar-10, Fashion-Mnist, and Mnist datasets, respectively.