Image Generating And Its Application With Wasserstein Generative Recurrent Adversarial Networks

Most image processing methods use supervised algorithms,while supervised learning has higher demand for image data and requires a large number of labels,which leads to a sharp rise in labor costs.Unsupervised learning,however,reduces the manual marking requirements for image data.The generative model is a typical unsupervised learning algorithm,and it is also a typical deep learning algorithm with good semantic understanding ability.The generative model can learn the distribution of real data and can sample realistic samples from it.Generative adversarial networks(GAN)is one of the mainstream generative models.Using game theory,a learning model and a discriminative model are constructed using deep learning.They are used for confrontation learning,and they continuously improve their ability through mutual feedback.,and ultimately generate high-quality data samples.This paper is based on the basic principle of generative adversarial network model,aiming at the problem of training instability in the original GAN,and simulating the process of human painting,improving Generative recurrent adversarial networks(GRAN)and proposing the Wasserstein generative recurrent adversarial networks(WGRAN).The main innovations and research results of this paper are as follows:(1)Simulate the process of the artist's repeated iterative and multiple modifications while painting,propose a recurrent generation model,and use the “multiple generation” method to generate samples.The generated sample at time t is added to the generated sample at time t-1,and the generated sample at the final time is taken as the output of the generator.This is different from GRAN's superimposing of generated samples at all times as the final generated sample,our method not only reduces the amount of calculations by the generator,but also does not degrade the quality of the generated sample.(2)GRAN model adopts the min-max objective function in the original GAN.However,due to the irrational distance metrics in the objective function,the model is unstable,which results in poor or poorly generated sample diversity.Therefore,this paper chooses Wasserstein distance as a new distance measurement method,adopts the weighted clipping technique to constrain the discriminant ability of the discriminator,and proposes the WGRAN model.(3)For the problem that the WGRAN model is difficult to optimize due to the weight clipping technique,the gradient penalty is used instead of the weight clipping to fine tune the objective function,and WGRAN's improved model WGRAN-GP is proposed.(4)This paper conducts experiments on four different types of datasets,and evaluates the quality and diversity of the generated samples using the two evaluation criteria of generative adversarial metric and inception score.Through the analysis of the experimental results,the model proposed in this paper has obtained good results on the relevant evaluation criteria.From the visual results of the sample generated by the generator,the quality of the generated sample is higher.