Latent Space Distribution Learning In Generative Adversarial Networks And Application

With the development of deep learning,the computer's cognition of things is no longer limited to the classification,identification and positioning of data,and at the same time,it can realize data synthesis,which is a high-dimensional understanding of data.Archaism cloud: knows what it is,knows why it is.When data can be modeled in high dimensions,the data will rise to a new stage on the cognitive level.Variational autoencoder(VAE)has taken the first step in the field of data generation;Stream-based Glow solves the problem of data inverse mapping;The game update of the generative adversarial network(GAN)discriminator and generator enables the model to realize the modeling from scratch in the data fitting,which not only displays the powerful ability in the synthesis of low-dimensional data,but also in the generation of highdimensional data.However,the generation of VAE is ambiguous,the huge calculation amount of Glow becomes a limiting factor for development,and GAN has the problem with unstable training and insufficient diversity of generated data during the data generation stage.How to achieve the balance and excellent performance of the generated model has attracted much attention.It is worth noting that the input of the GAN is random noise,which makes the generator completely achieve the task out of nothing.While VAE is from the coding of data,and then generates the data after obtaining the latent space of the data.How to use the low-dimensional latent space containing data hidden information to guide generative adversarial network is a key point to achieve high-quality data synthesis.This paper learns the latent space distribution of arbitrary data and feeds it into the generator to guide the generation of realistic data.This not only implements the stability of the GAN during the training phase,but also improves the problem of insufficient data diversity.At the same time,the experimental quality of the data is compared with the advanced baseline models to verify the advanced nature of the model.It has been applied to more tasks and its advantages have been demonstrated in image translation tasks.Specifically,the main research results of this paper are as follows:1.The learning algorithm of arbitrary data latent space distribution is designed.The low-dimensional latent space distribution is obtained by random noise fitting data coding,which provides the latent space variable with hidden information for the generation of high-dimensional data.2.The learned latent space variables are combined with a generative adversarial network.The closed-loop generative model algorithm CAE-CGAN is designed to achieve high-quality data generation with supervision and unsupervision.The qualitative and quantitative experimental results have been verified.3.Expanding the application of CAE-CGAN has made improvements under the task of image translation.Meanwhile,CAE-CGAN can be extended to general data conversion tasks.