DMGAN: Discriminative Metric-based Generative Adversarial Networks

摘要

With the proposed of Generative Adversarial Networks (GANs), the generative adversarial models have been extensively studied in recent years. Although probability-based methods have achieved remarkable results in image synthesis tasks, there are still some unsolved challenges that are difficult to overcome. In this paper, we propose a novel model, called Discriminative Metric-based Generative Adversarial Networks (DMGANs), for generating real-like samples from the perspective of deep metric learning. To be specific, the generator is trained to generate realistic samples by reducing the distance between real and generated samples. Instead of outputting probability, the discriminator in our model is conducted as a feature extractor, which is well constrained by introducing a combination of identity preserving loss and discriminative loss. Meanwhile, to reduce the identity preserving loss, we calculate the distance between samples and their corresponding center and update these centers during training to improve the stability of our model. In addition, a data-dependent strategy of weight adaption is proposed to further improve the quality of generated samples. Experiments on several datasets illustrate the potential of our model.