Sample Generation Combining Generative Adversarial Networks and Residual Dense Networks

Abstract

Recently, the generation of adversarial networks has made great progress in image generation and image enhancement, and it is even able to generate high-quality false images to deceive the human eyes, but Generative Adversarial Networks (GANs) still have problems, such as training process instability and mode collapse. To solve the problems above, we use the dense residual network and the residual networks to construct a generator and a discriminator of the networks Combing the Generative Adversarial Networks and Residual Dense Networks (RDGAN), respectively, and use the spectrum normalization model to constrain the GAN networks which can prevent the parameter size. To avoid gradient anomaly, combining with the TTUR optimization strategy, we design and implement several simulation experiments on the 102 Category Flower Dataset. Experimental results show that our method is superior to most existing methods in most cases.