Research On Unsupervised Domain Adaptation Algorithm Based On Adversarial Method

Image classification is one of the classical problems in computer vision.With the booming growth of data scales,there are huge challenges in image classification of unsupervised scenarios.Unsupervised domain adaptation,an effective method for unsupervised image classification,has obtained extensive and in-depth research in recent years.In unsupervised domain adaptation,the problem of knowledge learning in the target domain of unsupervised scenarios is mainly solved by transferring knowledge from a large amount of labeled images in related source domain.There are different data distributions between source domain and target domain,and the cross-domain gap is the main obstacle of knowledge transfer in unsupervised domain adaptation.In recent years,generative adversarial networks,an effective method for data distribution matching and generating,have been widely studied.As a consequence,a large number of researchers have introduced adversarial learning into unsupervised domain adaptation to solve the problem of cross-domain gap.At present,the domain adaptation algorithims still has the following problems: Firstly,the existing research does not fully consider the robust decision boundary of classifier,and the domain adaptive classifier cannot migrate to the target domain as well as expected,which reduces the classification accuracy of images in the target domain.Secondly,the existing research,which mainly considers the knowledge transfer from the source domain to the target domain,ignore the implicit feedback knowledge of images in the target domain,resulting in insufficient knowledge transfer in the source domain.Finally,the closed set domain adaptation algorithms existing negative tranfer problem in the scenario of unsupervised partial domain adaptation,and the researches on partial domain adaptation are insufficient.To solve the above-mentioned problems,the main work of this thesis are as follows:1)Investigate and study scientific research literature in related fields such as semisupervised learning and generative adversarial networks.Analysis and research on existing unsupervised domain adaptation algorithms and unsupervised partial domain adaptation algorithms.2)A framework of unsupervised domain adaptation algorithm based on consistency regularization is proposed(Consistency Regularization Domain Adaptation,CRDA).In the existing research,it is not easy to learn a classifier with robust decision boundary for images in the target domain.Inspired by consistency regularization in semi-supervised learning,a framework of unsupervised domain adaptation algorithm that is consistency regularized for images classification in the target domain is proposed in this thesis,and thus improves the images classification accuracy of source domain classifier in the target domain.3)An unsupervised domain adaptation algorithm that resampling confusion samples in the target domain is proposed(Resampling Adaptation Network,ResAN).In the existing research,it is not easy to mine and exploit the images classification knowledge of target domain in unsupervised scenarios,resulting in the problem of insufficient knowledge transfer from source domain to target domain.To solve this problem,in this thesis,two source domain classifiers are adapted to achieve domain adaptation.In order to effectively mine the knowledge of images classification in the target domain,an unsupervised domain adaptation algorithm that based on adversarial learning and resampling confusion samples in the target domain is proposed,which could promote positive transfer from source domain to target domain,and improve the classification accuracy of images in the target domain.4)An unsupervised partial domain adaptation algorithm that is adaptive knowledge transfer is proposed(Adaptive Knowledge Transfer Network,AKTN).In the scenarios of partial domain adaptation,the research on partial domain adapation algorithm is insufficient,and the classical closed set domain adaptation algorithms have the problem of negative transfer.To solve this problem,an adaptive knowledge transfer algorithm for unsupervised partial domain adaptation is proposed,which could promote positive transfer of shared classes images and reduce negative transfer of unshared classes images between source domain and target domain,achieving out-performance classification accuracy of images in the target domain.5)For the proposed algorithms,extensive experiments and analyses are performed on multiple public datasets.The experimental results demonstrate that the proposed algorithms achieves better classification performances of images in the target domain,which verifies the effectiveness of the proposed algorithms.