Research On Domain Adaptation Classification Under The Generative Adversarial Network

Generative Adversarial Network(GAN)is a new framework that uses adversarial processes to estimate generative models.The goal of GAN is to study the collection of training samples and learn to generate the probability distribution of training samples.Generate more samples from the estimated probability distribution.As one of the most successful generative models,GAN has attracted the attention of many researchers since it was proposed.Therefore,many GAN-based derivative models have appeared.Deep Convolutional Generative Adversarial Network(DCGAN)is a GAN-based derivative that introduces the Convolutional Neural Network(CNN)into the traditional GAN model for unsupervised confrontation training,but DCGAN still has greater difficulties in learning deeper features,which affects the effect of image generation.In recent years,GAN has also been used in many fields of adaptive classification tasks using its unique confrontational ideas.In traditional machine learning algorithms,it is usually assumed that the training samples and test samples come from the same probability distribution.But in fact,in many current learning scenarios,the probability distribution of training samples and the probability distribution of test samples are different.Domain Adaptation method,as a technique that can continuously reduce the difference between samples with different probability distributions,has attracted widespread attention in the field of Transfer Learning.Therefore,it can be explored in depth.In response to the above problems,this paper conducts an in-depth analysis from the three directions of generative adversarial network,single-source domain adaptation,and multi-source domain adaptation.The main work carried out is as follows:Firstly,Res DCGAN is proposed,and the idea of residual network is introduced into DCGAN to realize the complete transmission of feature information and retain effective parameters.It can study the characteristic information between the generated data and the real data,and use the quantitative evaluation method Frechet Inception Distance(FID)as a measure of the quality of image generation.The experimental results on three benchmark datasets show that the image generation effect achieved by this method is further improved with the decrease of the FID value.Secondly,SDAGAN is proposed.On the basis of Res DCGAN,the domain adaptive methods is adopted to improve the performance of image classification.Conditional Maximum Mean discrepancy(CMMD)is used to reduce the conditional probability distribution distance between source domain and target domain.The t-SNE visualization method is adopted as a criterion to evaluate the variation of conditional distribution differences among different domains.Seven mainstream methods are compared on three benchmark datasets.The results show that the classification accuracy of the SDAGAN is improved.Lastly,for the limited features of single-source domain learning,the SDAGAN model is further extended to multi-source domain adaptation,and the MSDACG is proposed.In this model,the effective features of multiple source domains are extracted.The domain-specific GAN structure and CMMD distance evaluation criteria are used to narrow the gap between different source domains and targets.After the introduction of multiple sources,the classifiers trained in different source domains are different.Therefore,the distance constraint between different classifiers is combined with the difference loss.Compared with the current nine mainstream multi-source domain adaptive methods on five different datasets,the image generation effect and classification accuracy have been significantly improved.