Research On Passive And Unsupervised Domain Adaptation Method Using Implicit Source Information Minin

To address the problem of domain shift caused by different distribution discrepancies between labeled source domain and unlabeled target domain,Unsupervised Domain Adaptation(UDA)that utilizes cross-domain knowledge obtained from labeled source domain data to assist in discriminating learning in the unlabeled target domain has been proposed.However,the source domain data accessible in UDA does not always be allowed to access directly for realistic reasons such as privacy protection,data storage,and transmission costs.Therefore,to address the challenge of restricted access to source domain data in practical UDA applications,Sourcefree Unsupervised Domain Adaptation(SFUDA)has received widespread attention,which extracts valuable knowledge for target domain adaptation from pre-trained models in the source domain without directly accessing source domain data.Although much current research has explored the SFUDA problem,there is still a lack of effective utilization of implicit source information.Therefore,the thesis delves deeper into the implicit source information at the sample prototype and model structure levels and effectively embeds it into the source-free unsupervised domain adaptation model for knowledge transfer.In summary,the main contributions of this thesis are as follows:1)Mining and modeling the implicit source information of the inherent sample prototypes.The current SFUDA methods for generating data replace inaccessible source domain data by generating pseudo-source data but fail to consider the implicit source information of the inherent sample prototypes in the target domain.Although the sample transformation method uses the target domain to generate pseudo-source data,this generation method does not have stability and flexibility.Therefore,this thesis proposes an SFUDA with Implicit Alignment of Trusted Pseudo Samples(SFUDA-TPS),which solves the SFUDA problem by constructing reliable feature prototypes using trusted pseudo-samples.Specifically,SFUDA-TPS constructs a target feature correction classifier to alleviate the problem of feature prototypes,in which the source domain pre-trained model estimates deviate from the target domain sample distribution.Based on this,the implicit source information of the inherent sample prototypes are mined by selecting the target domain samples with high information content as trusted pseudo-samples and then estimating reliable feature prototypes.Finally,the implicit alignment of the source and target domain is achieved by learning the source domain information hidden in the fixed source domain classifier.Experiments on several benchmark datasets demonstrate the effectiveness of the proposed method on SFUDA tasks.2)Mining and modeling the implicit source information of the deep structure of the model.Most existing source-free unsupervised domain adaptation methods perform source domain knowledge transfer by target domain pseudo-labels obtained from source pre-trained model.However,the source pre-trained model is a fixed model biased towards the source domain and away from the target domain.Therefore,the target domain pseudo-labels obtained by the current SFUDA method cannot remove the wrong pseudo-labels.Using these pseudo-labels will cause cumulative errors in the model and thus degrades the model performance.Therefore,this thesis proposes an SFUDA with maintaining Model Balance and Diversity Mining(SFMBD)to avoid using pseudo-labels with certain errors to train the target model.Firstly,the implicit source information in the source pre-trained model is mined by constructing a target domainspecific classifier with a decision boundary removed from the high-density region of the target domain feature distribution and transferred to the target model.Secondly,the proposed method promotes the model to be discriminative to the target domain while maintaining the balance and diversity of the model further to explore the deep structural information of the model.Finally,the experimental evaluations illustrate the effectiveness of the proposed method in solving SFUDA tasks.