Deep Feature Similarity for Generative Adversarial Networks

Abstract

We propose a new way to train generative adversarial networks (GANs) based on pretrained deep convolutional neural network (CNN). Instead of directly using the generated images and the real images in pixel space, the corresponding deep features extracted from pretrained networks are used to train the generator and discriminator. We enforce the deep feature similarity of the generated and real images to stabilize the training and generate more natural visual images. Testing on face and flower image dataset, we show that the generated samples are clearer and have higher visual quality than traditional GANs. The human evaluation demonstrates that humans cannot easily distinguish the fake from real face images.