Abstract: The emergence of generative adversarial network (GAN) plays a great role in solving the problem of insufficient sample data in the field of deep learning. In order to solve the detail quality problems of images generated by GAN such as foreground and background separation and contour blurring, this paper proposes an improved Wasserstein generative adversarial network with gradient penalty (PSWGAN-GP) method. Based on the Wasserstein distance loss and gradient penalty of WGAN-GP, this method uses the features extracted from the three pooling layers of the VGG-16 network in the discriminator and calculates the style-loss and perceptual-loss from these features as penalty terms of the original loss, which improves the discriminator’s ability to acquire and discriminate deep features and enhance the details of the generated images. The experimental results show that PSWGAN-GP can effectively improve the quality of generated images with the same generator and discriminator network structure and the same hyperparameters, and the scores in IS and FID are improved relative to other image generation methods.

Key words: deep learning, Wasserstein generative adversarial network with gradient penalty, VGG-16