Research On Conditional Generative Adversarial Networks Model Based On VAE

In recent years,the wave of artificial intelligence,led by deep learning,has achieved very good results in the field of supervised learning.Unsupervised learning,as a way to truly enable computers to learn from real world data without labels,can avoid trivial and data labelling work which can't be avoid in supervised learning.The best way to make computers understand the complex real world is to let computers generate the representation of the real world in a certain way.The first thing to do to achieve this goal is the generation model.In recent years,the most prominent in the generation model is the Variational Autoencoder and the Generative Adversarial Networks introduced in this paper.As an extension of the automatic encoder,the former combines the depth learning idea with the statistical learning well.The highdimensional distribution of the image can be reduced by the encoder network,and then the decoder network can be used to automatically generate an image similar to the original image from the low-dimensional data distribution.The Generative adversarial networks utilizes the zero-sum game theory,and alternately trains with generator and discriminator.Finally,it can achieve the operation from random noise to real image generation.However,although the variational autoencoder realizes the approximate output of the real image,the image generated by it is always fuzzy and could not generate high resolution images.For generative adversarial networks,its training is very difficult,and sometimes the generated image will be distorted seriously,which will lead to meaningless of image generation,and even the phenomenon of model collapse.Therefore,in the later improvement,the researchers took advantage of the combination of variational autoencoder and generative adversarial networks,and trained jointly,and achieved good results.But the inherent shortcomings of the original generative adversarial networks lead to the combination still can not achieve a very good effect.Based on this,in the paper,we propose utilize the idea of Conditions Generative Adversarial Networks—the improved based on the generative adversarial networks,and combine the CGAN to try to add artificial control in the process of image generation to improve the generative adversarial networks based on the variational auto-encoder.At the same time,the use of deep convolution generation to combat the improvement of the network structure and the WGAN idea of the network,and the introduction of gradient penalty method to optimize the generational confrontation network based on the variational autoencoder.In the algorithm flow of the generation network based on the variational autoencoder,we use the image and the condition feature map as the input to improve the algorithm to achieve the artificial control of the image generation process.At the same time,the deep convolution neural network is introduced to optimize the generative adversarial networks,so that GANs can get better convergence speed and model stability.By combining variational autoencoder with the improved GANs,the features are extracted by the variational autoencoder,and the extracted feature vector and conditional feature graph are used as the input of the generative adversarial networks,and the advantages of the two in feature extraction and generation ability are used respectively.Good effect.Moreover,we use the condition to generate the characteristics of the confrontation network,and realize the artificial control of the image generation process,then the new model is proposed and named VAE/CGAN.In order to verify the feasibility of the proposed model,we have carried out a lot of experiments on the CelebA face image data set.Compared with the original VAE,GAN and the improved VAE/GAN,the experimental results have achieved better results in the image clarity,model stability and so on.