Learning-aware feature denoising discriminator

Abstract

Although generative adversarial networks (GANs) show great prospects for the task of image synthesis, the quality of synthesized images by existing GANs is sometimes inferior to real images because their discriminators cannot effectively learn robust identification features from input images. In addition, the training process of discriminator is prone to be unstable. To this end, inspired by the denoising auto-encoders, we propose a learning-aware feature denoising discriminator. It is designed to pay attention to robust features of input images, so as to improve its robustness in identifying features and recognition ability in training process. First, we use a decoder to generate perturbing noise and add it to real image to get corrupted image. Then, we get the encodings of the corrupted image and real image through an encoder. Finally, we minimize both types of encoding to constitute a denoising penalty and add it to the loss of the discriminator. We also show that our method is compatible with most existing GANs for three image synthesis tasks. Extensive experimental results show that compared with baseline models, our proposed method not only improves the quality of synthesized images, but also stabilizes the training process of discriminator.