| Traditional machine learning requires two assumptions: training data and test data are independently distributed,with a large number of labeled training data.In practice,the above two assumptions are not satisfied,and the performance of the model obtained by machine learning is difficult to guarantee.Domain adaptation breaks through the above two assumptions,which classifies label-free target domains by transfer label-rich source domain knowledge.However,most traditional domain adaptation methods usually assume equal label space in the source domain and the target domain,leading to a weak generalization ability of the model to the target data.The emergence of partial domain adaptation breaks this constraint and assumes that the target domain label space is a subset of the source domain label space,making the generalization ability of the model superior.Most of the existing partial domain adaptation methods only use a single way to distinguish the samples in the source domain,which leads to inadequate screening of abnormal class samples,thus affecting the model learning performance.To this end,this paper studies partial domain adaptation,constructs two different adversarial learning-based partial domain adaptive models,and applies the model to pattern recognition and medical diagnosis scenarios.The main work of this paper is as follows:First,we propose a partial adaptation model based on the two-channel weighting mechanism(BDC＿PDA).The model uses two-channel joint efforts to screen the sample source in the source domain and assign corresponding transferable weights to each source domain sample.In addition,the model chooses to use adversarial learning to game the feature extractor and discriminator through the loss function to realize cross-domain migration.To achieve partial domain adaptation,the corresponding transfer weights generated by the source domain sample are multiplied by the loss generated by transfer,dynamically adjusting the contribution of the sample in the adaptation process.The effectiveness of the proposed method is also validated on public datasets such as image＿CLEF,Office31,and Caltech-Office.Secondly,we propose a domain adaptation model based on multi-attention representation network(MARPDA).The model first extracts the initial feature representation using the pre-trained Resnet-50.To extract more fine-grained features in the medical field,we separately input the initial representations into multiple convolutional networks with attention mechanisms to learn multiple different representations.The learned multiple representations are then spliced as input in the next network.Multi-representation features will be input to the auxiliary classifier module to extract sample weights and assign the output weights to the classification loss and adversarial loss for partial transfer.We apply the model to the X-ray image diagnostic problem of Covid-19 and perform relevant simulation experiments that demonstrate the effectiveness of the proposed method. |
