| Unsupervised domain adaptation is an important part of transfer learning.It learns a target model by knowledge transfer from a related source domain,to solve the learning problem of the task domain without labels,and commonly assumes the direct access to source data.Due to privacy or security concerns,only source model is available in some real applications.Source-free domain adaptation is a new learning scenario that only uses pre-trained source models to adapt to the target domain without source data.This approach has important research significance and application value.How to effectively utilize pre-trained model and target data to alleviate domain shift is a huge learning challenge of source-free domain adaptation.Therefore,this paper has conducted in-depth research on this issue,including the following two aspects:On the one hand,source-free domain adaptation with Dual-correction Mechanism(DCM)is proposed.Firstly,DCM attempts to exploit the target data structure to refine the noise-like prediction.Secondly,the teacher-student model is used to guide the learning of features,to maximize the consistency between high-confidence features and the difference between low-confidence features.Finally,experiments on Digit,Office-31 and Office-Home datasets show the effectiveness of DCM.On the other hand,source-free domain adaptation with Generated Target-like Prototypes(GTP)is proposed.Firstly,to make use of source domain classifier knowledge and target data features,a conditional generative adversarial network is used to generate a target-like prototype.Secondly,GTP calculates the distance from the target sample feature to the prototype to obtain the pseudo label and fine-tunes the target model under this guidance.Finally,significant learning performance improvements have been achieved on Office-31,Office-Home and Vis DA datasets. |
