Research On Domain Adaptation Algorithm Based On Metric Learning

In recent years,deep learning has made breakthroughs in many fields,and its excellent performance depends on supervised learning trained with a large amount of labeled data.Since tagging images is very labor-intensive and time-consuming,domain adaptation methods are proposed to apply the trained deep model of the source domain to the target domain without label information..However,there exists a domain shift between the source domain and the target domain.In order to address this issue,metric learning is introduced to improve the unsupervised domain adaptation for enhancing the classification accuracy in the target domain,which includes:(1)In order to improve the classification accuracy of the unlabeled target domain,an unsupervised domain adaptation algorithm using angular margin loss is proposed.Specifically,the angular margin loss function is utilized to reduce the angles between the sample feature vector and the weight vector by metric learning in tha angle space.In such case,the better clustering effect of the source domina can be obtained.Additionally,the domain alignment is performed by the domain adversarial training.Hence,the intra-class distance of the target domain can be reduced indirectly.Therefore,the classification accuracy on the target domain can be significantly improved due to strong generalization ability.(2)In order to improve the classification accuracy of confusing classes in the target domain,an unsupervised domain adaptive algorithm using triplet loss is proposed.First,the boundary threshold is calculated by the classification distance in the target domain to identify the confusing class,and the confusing categories in the source domain are separated by dynamically adjusting the margin in the triplet loss.Furthermore,the confusing categories in the target domain are also separated indirectly by domain alignment,which will result in the further improvement in classification accuracy in the target domain.In addition,an effective training batch generation mechanism is employed to accelerate the training process.Finally,multiple experiments were conducted on some domain adaptation benchmarks.The experimental results show that our proposed methods outperform various compared methods,which further verifies the feasibility and advantage of the proposed methods.