Research On Image Generation Technology Based On GAN And VAE Fusion Network

With the continuous development of computer vision technology,image generation technology has gradually become an important research topic in the field of computer vision.Natural image generation models such as Variational Auto-Encoder(VAE)and Generative Adversarial Network(GAN)are the most familiar technologies in this field.VAE introduces implicit variable inference and log likelihood estimation,the training process is stable,but the generated image is lack of detail description,and the contour is fuzzy;the image generated by Gan is real and clear,but the training process is prone to mode collapse,which is more serious in the synthesis of high-resolution image.In view of the above problems,this paper studies VAE and Gan from the following three aspects.1.To improve the image generation technology,two kinds of image generation models,VAE and Gan,are fused.The encoder,generator and discriminator of VAE are fused in series.Tandem VAE-GAN(TVAE-GAN)based on GAN and VAE is proposed.VAE is used to encode the input image,and then the extracted coding features are transferred to Gan as input instead of the traditional noise.Through the feedback information of Gan discriminator,the quality of the generated image is continuously improved,and finally the generated image with high quality is obtained.The experimental results show that the image generated by TVAE-GAN is clearer.2.In view of the complex structure of TVAE-GAN network,the time cost of image generation increases,and the convergence speed of the algorithm slows down,mixed and fused GAN's generator with VAE's coder,replaced the discriminator in TVAE-GAN with the encoder and eigenvalue function,and put forward the hybrid Mixed VAE-GAN(MVAEGAN)network based on both VAE and GAN by adding similarity measurement function.The results of image generation in MNIST handwritten digital data set and CelebA face image data set showed that MVAE-GAN had higher image generation quality,reduced time overhead and accelerated algorithm convergence speed compared with TVAE-GAN.3.In order to further verify the practical application value of MVAE-GAN in image generation,MVAE-GAN was applied to the removal of face mosaics and achieved a good removal effect.The research results were helpful to solve the problems of image acquisitionand data defects and meet people's demand for specific image data.