Based On MCMC Algorithm Improvements To The Quality Of Images Generated By Conditional Generative Adversarial Networks

With the deepening of machine learning and artificial intelligence research,image generation technology has attracted more and more attention.Generative Adversarial Network(GANs)provides a generation model by means of generator and discriminator confrontation.On the other hand,Conditional Generation Antagonistic Network(CGANs)inputs labels into the network on the basis of generating the antagonistic network,so that the network can generate pictures or results under the specified labels.Different construction methods of discriminator network also lead to different forms of CGANs.The appearance of CGANs provides a new idea for people to generate a certain kind of pictures,fit the data distribution and enhance the data set according to the conditional information.However,in practice,due to the minimax form of the GANs objective function and the characteristics of the measured JSD distance,the generated pictures often have poor sample quality,mainly including unclear pictures and poor diversity.Similarly,CGANs will have such problems.Therefore,this paper puts forward the Markov Chain Monte Carlo Algorithm(MCMC)to further improve CGANs and improve the quality of generated samples.In this paper,the Metropolis-Hastings(MH)algorithm in MCMC algorithm is used to take the real data distribution as the target distribution,and the two problems of acceptance rate and suggestion distribution are solved respectively.When calculating the acceptance rate,firstly,the gap between the real distribution and the generated distribution of data is deduced through the objective function of CGANs,and the gap,that is,the density ratio of the two distributions,is expressed by the score of the discriminator.Secondly,the Markov chain in the high-dimensional sample space is transformed into the low-dimensional potential space,and an easy-to-calculate acceptance rate is proved theoretically.When selecting the suggestion distribution,the suggestion distribution selects independent suggestion distribution and dependent suggestion distribution respectively.Finally,in the simulation experiment,for three labeled data sets:MNIST data set(labeled as numbers),CelebA data set(labeled as text)and Facades data set(labeled as pictures),it is verified that the quality of the generated pictures of MCMC-CGANs is better than that of CGANs,which provides a better method for generating pictures according to labels in the future.The innovation of this paper lies in the combination of Markov chain Monte Carlo algorithm and conditional generation confrontation network.When CGANs is used to generate samples,the samples are further screened,so that CGANs can generate samples with better quality.Three common forms of CGAN,ACGAN and Pix To Pix are explained in detail respectively.The method in this paper is as follows:firstly,the potential variables are sampled from the suggested distribution in the low-dimensional potential space of the generator,and the obtained suggested potential variables and labels are sent to the generator at the same time to generate corresponding suggested samples,and then the acceptance rate is used to calculate whether to accept this sample.This process is repeated for k steps,and finally the obtained samples will be closer to the distribution of the samples we want.