Research On Domain Adaptation Method Based On Deep Learning

At present,deep learning has achieved great success in various fields,such as image recognition,semantic segmentation,target detection,natural language processing and other tasks,but the premise of success in these tasks is to have a large number of data labels.However,in practical applications,some tasks are difficult to obtain data labels that meet the requirements,such as medical images,remote sensing images and other issues.So,we usually look for the data set of similar domain(source domain)to label,and then apply it to the original problem(target domain),but when there are differences between the source domain and the target domain,we can not get the predicted effect.Therefore,how to solve the deviation between the target domain and the source domain has become an important research direction of machine learning.The main work of this paper includes the following points:1)Aiming at the problem that the training target of the classic method DANN can only make the features of the source domain and the target domain overlap on the whole,but ignore the difference between different data and different categories,this paper proposes an unsupervised domain adaptive method based on label propagation.The purpose of label propagation is to make the source domain and the target domain infer each other's labels,and then construct cyclic consistency loss,which makes the data characteristics of the target domain and the source domain gather in the same label.2)The evaluation of the existing methods is only based on the transfer result from one source domain to another specified target domain,but there may be another new scenario in the application,for example,a network has been deployed in a system,which may need to switch back and forth between the source domain and the target domain,that is,the model cannot judge whether the samples to be tested come from the source domain or the target domain,but we still need to give the judgment.To solve this problem,we propose a domain adaptation method based on the residual correction module,which considers the balance of the algorithm in the source domain and the target domain,and analyzes and solves the problem about the maximum mean difference.3)Aiming at the scene that targert domain obtains a small number of labed samples,this paper proposed a semi supervised domain adaptive method based on active learning.The main difference between this method and unsupervised domain adaptive method is that the classifier obtained by learning on the source domain is better transferred to the target domain with a small amount of target domain label information.