Study Of "Intermediate State" On Generative Adversarial Network

In recent years,the generative adversarial networks(GAN)is a kind of image generation network,which is operated by two models of generative model and discriminative model through the method of adversarial learning.This idea of adversarial learning makes GAN stand out in the generative network model.In view of the problem that the output and input of the image generated by GAN are all of the same type,this paper proposes to construct a new type of GAN with different types of input and output.The main research is as follows:Firstly,the training process and principle of GAN are studied.Then,the structure and application of several derivative models of GAN are introduced,and the comparison between these derivative models and the original model is analyzed.Secondly,the paper analyzes the influence of the addition of discriminator network on the GAN by building SRGAN which can generate high-resolution image from low-resolution image.In addition,the function of generative model and the importance of discriminator network are proved by the separate generative model experiment.Combining with the optimal transmission theory to analyze generative adversarial networks.Finally,aiming at the problem that the input and output of GAN are mostly similar in the past,a method of generating"intermediate state" based on GAN is studied.This method is tested on the MNIST dataset and the handwritten Chinese character dataset.The experimental results show that the "intermediate state" based on GAN is a new thing which depends on the initial state and synthesizes the two main features of input,and the input with simple feature distribution is easier to be learned by machine.The work of this paper shows that the method of generating"intermediate state" based on GAN can use the generative adversarial network to generate "intermediate state" pictures,and construct a network that can explore "intermediate state" things.This method can be used to guide the process of finding new things.