Research On Image Generation Algorithm Based On Generative Adversarial Networks

As an important source of information,images are inseparable from people’s lives.The research on images has always been a key field of machine learning.Compared with image discrimination tasks,the task of generating images from existing data sets is more difficult.With the development of deep learning,the research of image generation has made great breakthroughs in content creation,intelligent editing,and data enhancement.More and more research teams use deep learning to deal with image generation related issues,Generative Adversarial Networks have also been proposed.In recent years,in order to expand the applicable scenarios of Generative Adversarial Networks,researchers have made many improvements to the model.This paper uses a variety of Adversarial nets to study the key issues and application scenarios in the image generation task.The main work includes the following parts:(1)Briefly summarize the development process of deep learning,understand a variety of network models based on deep learning,and clarify the working principles and application scenarios of these network models.Taking the example of image style transfer as the starting point of this article,it focuses on comparing the traditional style transfer algorithm based on non-parameter and the style transfer algorithm based on convolutional neural network,and the powerful ability of convolutional neural network in feature extraction is clarified through experiments.In addition,two typical neural style transfer algorithms are introduced in detail,including the comparison of network structure and loss function.After summarizing several style transfer algorithms,it can be seen that as the depth of the network increases,there will be too many model parameters,and the training process will slow down;the GAN depends on the selection of data sets,and the quality of the generated image is affected by the network structure and loss function.The main purpose of this part of the work is to provide theoretical basis and algorithm support for the study of image generation in this article.(2)Using depthwise separable convolution to improve the discriminator network structure of the Deep Convolution Generative Adversarial Nets(DCGAN).A part of the ordinary convolution in the discriminator network is improved to a depthwise separable convolution and applied to DCGAN image generation,and a set of oil painting style image generation is added.The experimental results show that the quality of the generated image is not reduced,and the entire network model only reduces the network parameters and the amount of calculation,which improves the calculation efficiency of the model.Depthwise separable convolution and ordinary convolution are equivalent to the results of model generation tasks,and provides technical support for the application of the generated image in the design field.(3)Under the influence of the multi-scale pyramid structure of the Single-image Generative Adversarial Networks(SinGAN),the generator network structure of DCGAN is improved and an image super-resolution reconstruction model based on multi-size generator is proposed.The original single network of the generator is transformed into a similar size pyramid structure to improve the resolution of the generated image.In addition,the loss function of content perception is introduced to reconstruct the highfrequency details of the image and reduce the loss of high-frequency information.Experimental results show that the super-resolution reconstruction effect of this method is better than that of similar models.