An Feature Image Generation Based on Adversarial Generation Network

摘要

Generative antagonistic network (GAN) was proposed in 2014 to assist in generating realistic visual images, which has become one of the most popular research objects in deep learning in recent years. In the field of image generation, GAN is more effective than the traditional method, but it is difficult to train, unstable network and difficult to convergence. In this paper, GAN is applied in the field of feature image generation, and a new framework is proposed based on the C-SEGAN. By adding additional condition features to generator and discriminator, the similarity of distributed error is learned, and the discriminator is self-encoder, the mean square error loss is added to discriminator, and the generated model generates the specified sample. The model can generate the specified clear image according to the feature conditions. The experimental results show that the method can achieve faster convergence rate and generate better quality and diversity images with a simpler network than other supervised class generation models.