Research On Multi-source Domain Deep Transfer Learning Method

In recent years,the Unsupervised Domain Adaptation(UDA)algorithm has become a research hotspot of deep transfer learning algorithms.The algorithm requires that only the source domain data has a label,and the target domain data has no label.At present,most UDA algorithms are based on single-source unsupervised domain adaptation research.However,in actual scenarios,the source domain data is often obtained from various sources,and the source domain data and target domain data obtained from different sources have characteristics The distribution is similar but the degree of similarity is different,and the data distribution of the source domain data obtained from different sources will also be different.The single-source unsupervised domain adaptation algorithm has shown inferiority in solving the actual situation of this multi-source domain.Excellent performance.Moreover,most of the current multi-source domain adaptation algorithms do not consider adding the difference information between the source domain and target domain feature distributions to the iterative training process of the network,instead,they use the feature information of each source domain to perform Transfer learning,this kind of processing cannot reasonably and effectively learn from the characteristic information of each source domain,which leads to the phenomenon of negative transfer.This paper focuses on the research of multi-source unsupervised domain adaptation algorithm.First of all,this paper considers the negative migration problem caused by the equalization of the feature information of each source domain in the process of multi-source domain adaptation.Based on the Wasserstein metric and the Maximum Mean Discrepancy(MMD)metric,a Multi-source UnBalanced Domain Adaptation(MUBDA)network is proposed;secondly,in order to further improve the recognition accuracy of target domain data,this paper proposes a sample optimization algorithm to avoid the migration of irrelevant information.(1)In the adaptation process of multiple source domains,the problem of negative transfer caused by the balanced utilization of the characteristic information of each source domain is addressed.First,based on the degree of difference between the feature distribution of each source domain and the target domain,this paper proposes a multi-source unbalanced domain adaptation adjustment mechanism;at the same time,this paper also proposes a single source weight adjustment mechanism based on the control variable method.Through experimental comparison,This further demonstrates the rationality of the multi-source unbalanced domain adaptation adjustment mechanism;secondly,this paper proposes the MUBDA network based on the multi-source unbalanced domain adaptation adjustment mechanism,which can draw on the characteristic information of each source domain unbalanced and weaken the impact of negative transfer;Finally,this paper evaluates the performance of MUBDA on multiple public unsupervised domain adaptation datasets,and compares it with the experimental results of multiple unsupervised domain adaptation methods,focusing on the comparison of the multi-source domain adaptation method DCTN The experimental results of M3SDA and MDAN verify the effectiveness of the MUBDA network.(2)Aiming at the problem of sample optimization in the process of multi-source domain adaptation.First,this paper constructs a single-source domain adaptation network based on the correlation alignment(CORAL)measurement;secondly,this paper proposes a sample optimization algorithm to solve the negative migration phenomenon based on the source domain data quality factors to avoid irrelevant information in the source domain.Migration;Finally,this article uses the sample optimization algorithm to eliminate negative samples by inputting the handwritten character data set into the single source domain adaptation network based on the CORAL metric,and improves the recognition accuracy of the target domain.At the same time,this article uses Office-31,Office-Caltech10 and ImageCLEF-DA datasets apply the sample optimization algorithm to the MUBDA network based on the Wasserstein metric,which further reflects the advantages of this article's sample optimization method in multi-source domain adaptation.The experimental results show that the method in this paper has achieved better recognition results.First,the MUBDA network proposed in this paper based on the Wasserstein metric has shown better recognition effects on the Office-31 dataset,Office-Caltech10 and ImageCLEF-DA dataset;secondly,this paper applies the sample optimization algorithm to the Wasserstein metric.In the MUBDA network,comparing the experimental results before and after sample optimization,the recognition accuracy of the target domain has been improved on multiple datasets.