PWGAN: wasserstein GANs with perceptual loss for mode collapse

Abstract

Generative adversarial network (GAN) plays an important part in image generation. It has great achievements trained on large scene data sets. However, for small scene data sets, we find that most of methods may lead to a mode collapse, which may repeatedly generate the same image with bad quality. To solve the problem, a novel Wasserstein Generative Adversarial Networks with perceptual loss function (PWGAN) is proposed in this paper. The proposed approach could be better to reflect the characteristics of the ground truth and the generated samples, and combining with the training adversarial loss, PWGAN can produce a perceptual realistic image. There are two benefits of PWGAN over state-of-the-art approaches on small scene data sets. First, PWGAN ensures the diversity of the generated samples, and basically solve mode collapse problem under the small scene data sets. Second, PWGAN enables the generator network quickly converge and improve training stability. Experimental results show that the images generated by PWGAN have achieved better quality in visual effect and stability than state-of-the-art approaches.