Research Of Image Generation Technology Based On DCGAN Algorithm

In recent years,Deep learning has received increasing attention from researchers.Compared with the traditional machine learning,the multi-layer network structure of the deep learning model can express the complex function more effectively,so as to learn the characteristics of the stronger characterization.Generating adversarial Nets as a branch of deep learning has gained wide attention since its emergence.GAN is of great significance to the development of generative model.It can effectively solve the problem of generating natural interpretation data as a method of generating method,especially for generating high-dimensional data,and the neural network structure used by GAN is not limited Dimension,greatly extending the scope of the generated data samples.In addition,GAN in the field of natural language processing also has effective ability,such as generate dialogue,text to images and so on.This ability to generate infinite samples,respectively,has significant application value in AI such as image and visual computing,speech and language processing,interconnection and large system information security.(1)This paper introduces the background and significance of this study,expounds the history of deep learning and GAN,and the research status at home and abroad.And the important knowledge points used in this paper include convolution,transpose,convolution,pooling,VGG-19 model,residual network,GAN model,long and short term memory model,etc..The focus of this paper is to combine these existing methods effectively,and achieve the purpose of generating images.(2)GAN with unlimited build capacity,and we hope that this generating capacity can be carried out in accordance with the needs of purpose.Based on this,a model of generating images according to language description is constructed.The LSTM algorithm is used to extract the linguistic features,and the generated features are used to construct the graphs.The repetitive training of the generator and the discriminator is used to improve the image quality.The VGG-19 is used to optimize the LSTM to improve the effect of the extracted language feature.(3)GAN's ability to generate is not only reflected in the infinite,but also reflected in the small to large.The high-resolution image generation method based on GAN and residual network is based on small image,and after multiple feature extraction of the residual network,the transpose convolution operation is expanded to generate a high-resolution image.The training of the generator and the discriminator improves the quality of the generated high-resolution image.