Unsupervised Domain Adaptation Research Based On Domain Relation Utilization

Domain adaptation(DA),as one of the important directions of machine learning,has achieved important application results in semantic segmentation and image detection.In domain adaptation,the target domain samples are completely unlabeled,which is called unsupervised domain adaptation(UDA).At present,although many UDA models have been proposed,most of these methods only consider how to use the source domain information,and do not dig the relationship information between domains from a deeper level.In addition,with the current domain adaptation scenarios becoming more and more severe,the information available in the source domain is decreasing.Therefore,the adaptive scenarios such as sourcefree domain adaptation and whole unsupervised domain adaptation are derived.Therefore,this paper attempts to explore the deeper relationship between source domain and target domain,and to model and express it.At the same time,it tries to explore deeper relationship in the domain with less supervision information and improve its modeling representation.In a word,the main work of this paper is as follows:1)Exploiting The balance between class weight and sample weight.In order to solve the problem that the class weights and sample weights do not mismatch between source and target domains,this paper proposes SCUDAN model.On the one hand,after the class weight is determined according to the sample number of each class,the weight alignment is realized by eliminating the class weight difference.On the other hand,the model quantitatively represents the spatial distance between the sample and the class center,and defines the weight of each sample by the space distance,and realizes the alignment of the sample weight.Finally,the model is improved by using the existing CNN network as the feature extractor.A large number of experiments show that the SCUDAN model achieves better results,more effective and advanced than the existing domain adaptation models.2)Exploring the deep relationship between classification models in source-free UDA scenes.To solve the problem that the classifiers in source-free UDA is not suitable on the target domain,this paper proposes the STDA algorithm,which first uses the target domain to construct the pseudo source domain for replacement.Then,a sample transfer decision function is designed to judge the confidence of the samples in the target domain.The samples with high confidence are regarded as reliable samples.Then,with the help of these reliable samples and the pseudo labels given by the classifier,the conditional distribution is aligned,and the trusted label as signment of the target domain samples is realized by the above operation.Through a large number of experiments,the effectiveness and performance superiority of STDA algorithm in the sourcefree domain adaptation scene are proved.3)Exploiting deep information under single source multi-objective whole UDA.The com-bination of whole UDA and single source multi-objective domain adaptation leads to more un-supervised information and less available supervised information.Therefore,with the help of dictionary learning,this paper proposes two domain adaptation models: UDA-SKTR model and DA-MTDA model.In the former,the dictionary and the source domain projection are used as the source domain information set available for supervision,and the sparse representation of the dictionary is used to make the target domain extract information from the source domain that conforms to its own domain characteristics.The latter uses a unified dictionary throughout the source domain and the target domain to store the shared knowledge among the domains.Moreover,considering the differences between the target domains,DA-MTDA enters the differential dictionary to describe the unique knowledge of the target domain itself,so as to complete the modeling.Through a large number of experiments,it is verified that there are a lot of implicit relationships between the unsupervised target domain and the source domain.Using these relationships and modeling can significantly improve the accuracy of domain adaptation.