Research On Domain Adaptation Algorithm Based On Dynamic Inter-Class Distance And Pseudo-Labeled Clustering

With the continuous development of deep learning,the deep model has dramatically improved in the application of classification and prediction at the expense of massive amounts of labeled data.In real-world scenarios,the annotation of massive data is very labor-intensive and time-consuming,which often hinders the classification ability of deep model.In recent years,domain adaptation algorithms have been widely applied to adjust the trained deep model to unlabeled target domain by the means of domain alignment.However,due to the easily-confused classes and poor clustering effects in the target domain,the performance of domain adaptive algorithms in actual application needs to be further improved.(1)In order to improve the classification accuracy of unlabeled samples in the target domain,we propose a domain adaptation based on dynamic inter-class distance between easily-confused classes.First,the easily-confused classes are identified in the target domain according to the output difference of the classifier.Then,additive angular margin loss is used to reduce the intra-class distance of the easily-confused classes in the source domain for better clustering effect.And by introducing the interclass loss,the inter-class distance of the easily-confused classes in the source domain is increased by widening the angle between the weight vectors corresponding to these classes.Finally,through the domain alignment operation,the center distance of the easily-confused classes in the target domain is indirectly separated,which helps to enhance the classification performance of target domain.(2)Considering the poor clustering effect of samples in target domain due to the unsupervised training,we propose a domain adaptation method based on the seudo-label clustering of target domain,where two deep model for dynamic inter-class distance between easily-confused classes are trained separately.Specifically,contrastive learning is achieved by the feature extraction of the deep model.By distinguishing the softness and hardness of the pseudo-labels of the target domain samples,the clustering effect of target domain can be further improved by adopting the improved cross-entropy loss function and triplet loss function based on the dynamic average-value model.By above methods,the clustering effect of target domain can be further enhanced by reducing intra-class distance and easily-confused target samples,Therefore,the classification accuracy of the target domain will be improved accordingly.Finally,multiple comparison experiments were conducted on several publicly available domain adaptation datasets.Results show that our proposed domain adaptationalgorithm based on dynamic inter-class distance and pseudo-labeled clustering of the target domain outperforms traditional domain adaptive algorithms in terms of classification accuracy in target domains,and these results further demonstrate the effectiveness and superiority of our proposed method.