The Research And Application Of Improved Generative Adversarial Networks Technology

Generative Adversarial Network(GAN)with the idea of game confrontation has become a popular generative model.Compared with the traditional generative model,GAN not only has a generative model but also has a discriminative model that can distinguish the authenticity of the data against it.While GAN is conducting image generation research,the generated data and real data are used as the input of the discriminator.This method i s a good substitute for traditional data enhancement and is beneficial to image recognition.However,there are many problems in GAN,such as unstable training and difficult convergence,which greatly affect its development.Therefore,it is of great signi ficance to study and improve GAN to solve the existing problems and use GAN for image recognition.After analyzing the development and research results of generative adversarial networks at home and abroad,the basic idea of GAN,the training process of GAN,and the objective function of GAN,in view of the problems that the traditional GAN generated samples are uncontrollable and unclear,the training is unstable,and the input feature extraction is insufficient,the following main works are carried in this thesis:Aiming at the instability of traditional GAN training and uncontrollable and unclear generated images,a self-attention residual conditional generative adversarial network(SA-Res-CGAN)method is proposed with a generative adversarial network and residual network as the core.The method uses improved residual blocks instead of traditional convolutional blocks as far as possible,introduces a self-attention mechanism to calculate the weight parameters of input features to improve image feature extraction accuracy,and introduces image label information as a condition to better guide image generation,and finally introduces a gradient penalty to well solve the situation that the adversarial training mechanism of GAN tends to lead to unstable gradients of generators and discriminators and scattering of training.Experimental validation on three different types of datasets shows that SA-Res-CGAN can generate higher-quality images.In response to the fact that some researchers focus on the use of generators and ignore the powerful feature extraction capability of discriminators.In this paper,the proposed SA-Res-CGAN discriminator,compared with the traditional GAN discriminator,performs not only true/false judgment of the input data but also pre-training for recognition and classification of the input data,which plays a role in data enhancement.The trained discriminator that removes the fully connected layer of true/false judgment is extracted to construct classifiers for image recognition,where adaptive average pooling is used in the network structure of the discriminator to stabilize the spatial feature variation of the input and avoid overfitting of the model.The SA-Res-CGAN classifier is applied to two different types of datasets for experimental validation and compared with some existing analysis methods.The results show the effectiveness of the model for image recognition and further demonstrate the powerful feature extraction capability of the GAN’s discriminator.