Discriminator-Quality Evaluation GAN

In existing generative adversarial networks (standard GAN and its variants), the discriminator is trained for recognizing the real data as positive while the generated data as negative. This kind of positive-negative classification criterion ignores the fact that the discriminator is a non-objective evaluator, which means that the image quality evaluated by the discriminator may fluctuate during the whole training progress. Considering this fact, we propose a novel GAN framework called Discriminator-Quality Evaluation GAN (DQE-GAN) by using the discriminator outputs to evaluate image quality. By dynamically classifying images into high discriminator-quality and low discriminator-quality samples, every adversarial iteration step can be more reasonable and objective. The convergence of DQE-GAN framework can be theoretically proved. Through extensive experiments, we demonstrate DQE-GANs’ ability of achieving better generated images faster and more stable.

I. Introduction

Generative adversarial network (GAN) [1], as one of the most important deep generative models, can generate nearly realistic images under the guidance of min-max game training strategy. After years of development, it has been widely used in many fields such as image generation [2], [3], super-resolution image reconstruction [4]–[6], facial age synthesis [7], [8], and style transfer [9], [10]. The basic idea of GAN is to take noise vectors which come from a prior distribution as the input and make the generated data gradually approach the real image distribution through the competition game between a generator and a discriminator . The generator tends to produce data that follows the distribution of the reference natural images, while the discriminator aims to distinguish the generated data from the real data.