Virtual Image Generation Based On Generative Adversarial Networks

The research of image generation is of great significance in image field,such as style conversion,image super-resolution reconstruction,Natural language generation,medical data generation,physical modeling and so on.The most important thing in image generation is to choose the appropriate generation model,and a good generation model needs to be able to express the real data distribution.But the traditional method of learning data distribution function can only deal with some images with simple texture and regular structure.Therefore,how to generate complex natural images,and as far as possible to ensure the authenticity of the generated images,diversity has become an important issue in the field of deep learning.In this paper,we introduce the improved model of generative countermeasure network from the aspects of network structure and model loss function.In order to solve these problems,on the one hand,from the point of view of the structure of the Generative Confrontation Network Model,in this part,in the original GAN model,the fully connected network layer was replaced by the convolutional neural network layer to extract image features better,and the volume normalization method was used in the convolution network,the output of the feature layer is normalized to accelerate the training speed and improve the training stability of the model.In addition,Leak ReLU activation function is used instead of Relu activation function in order to improve the gradient disappearance of the original GAN model.On the other hand,in this part,in order to solve the instability problem of the original GAN in training,we improve the loss function of the generative countermeasure network model,the Wasserstein distance is used instead of JS divergence to measure the difference between the distribution of real data and the distribution of generated data.The smoothness of Wasserstein distance can deepen the gradient required by the model in training and effectively solve the problem of gradient vanishing,in order to balance the training level of generator and discriminator,the model parameters are pruned,and the RMSProp optimization algorithm which is suitable for the gradient instability of function is used to replace the kinetic energy optimization algorithm.Finally through these two aspects to improve the quality of the generated image model.We selected two classical data sets for empirical analysis and used IS and FID to measure the quality of different generation models.The final FID value of CIFAR-10 Dataset is 37.07 on the optimized GAN model,which is 61.39 lower than that of GAN model,and 27.37 FID on the optimized GAN model,this is 71.09 down from the original GAN model.Through experimental analysis,we can draw conclusions from the quality of the generated image: First,using convolutional neural network instead of full connection network in the original GAN model can extract image features better.The batch normalization method used in convolution network can effectively accelerate the training and improve the stability of the model training.Secondly,Wasserstein distance is used instead of JS divergence to measure the difference between the distribution of the generated data and the real data,and the training parameters are pruned with weights to deepen the training gradient of the model,effective prevention of gradient loss and mode collapse.Thirdly,the image quality and diversity of the modified GAN model are better than those of the original GAN model both in structure optimization and loss function optimization.