Design And Implementation Of Generative Adversarial Network Image Translation

Image is still the main way for people to communicate on social platforms,but the image is easily affected by the generation equipment and the way of transmission in the process of dissemination,causing pixel damage,so the research on images has always been an important field of computer vision.Generative adversarial networks has been a hot topic of machine learning since its introduction in 2014.The most obvious advantage of the generative adversarial network is that it is an iterative generative neural network model,which mainly includes two neural networks,namely the generator model and the discriminator model.The advantage of the generator is that it is easy to generate more deep models.The disadvantage is that the purpose of the generator is to imitate a certain target to get a high score,so the generator only learns the method to improve the score,and cannot improve the stability of the entire network.In this subject,the vae is selected as the generator of the generative confrontation network for experiments.Variational autoencoder is a generative model of deep learning model based on variational inference.In this paper,the variational autoencoder is used as the generator of the generative adversarial network to enhance the stability of the generative adversarial network training,and solves the problem that the generator cannot obey the standard deviation directional generation during the generation process,resulting in the loss of pixel information in the translation process and the compression of image quality The problem.This paper proposes a new Vae base GAN structure to improve the ultra-high-definition image quality of image translation.Two sets of control experiments were carried out in this subject.In the first experiment,VAE was selected as the generator model of the experiment for mnist generation experiment.At the same time,DCGAN is selected for comparison MNIST generation experiments under the same experimental conditions,among which the DCGAN model is mostly a generation model using a convolutional neural network.The model architecture optimization is divided into three parts to improve the encoder for encoding the original image,in order to generate the mean value ? and the standard deviation ? of the two parameters;use the standard deviation and the mean value to construct the hidden variables of the initial input coding of the image;use the hidden variables to pass decoding Reconstruct the image.Experiment two is an image translation experiment based on the pix2 pix network based on the variational autoencoder and the pix2 pix network based on U-NET.Pix2 pix GAN is essentially an image translation model,which translates one image into another.The input image can be regarded as a condition,and the transformation effect can be obtained according to the condition after the given condition.The generator model uses the vae model,and uses skip-connection to share more information.When generating images,the model can accept both high-level information and low-level information that is not lost as output.Adding Patch-D to optimize the discriminator model makes the discriminator a fully convolutional network,discarding the fully connected layer,and obtaining the output range of the discriminator through a part of the receptive field.In order to ensure the corresponding relationship,the loss function L1 is added,and the high-level abstract features of the generated model image and the original image are more similar.The experimental control group pix2 pix uses the original U-NET as the generator model G.Compared with the network proposed in this paper,he training speed is slower,the generated image loses the transmitted content,and the generation effect is not clear.