Unsupervised Adversarial Domain Adaptation Method And Application

The inconsistent distribution of data and the lack of sample labels have always restricted the performance of deep learning models.While domain adaptation breaks the distribution consistency assumption,knowledge transfer is performed to solve the problem of unlabeled or few labeled data.In this paper,we mainly focus on the unsupervised domain adaptation.And we mainly study image pattern recognition and medical diagnosis by building an adversarial unsupervised domain adaptation framework based on transfer learning.The main content of this paper is as follows.Firstly,an unsupervised adversarial domain adaptation model(NMADA)based on noise residual mixup is proposed to solve the problem that the inter-domain information in the shared latent space between two domains cannot be fully considered in UDA methods based on deep adversarial learning and some low-level feature information is usually lost after multiple convolutions.In NMADA,noise residual module receives low-level features of the source domain and generates new source domain features.Mixup convex optimization is performed on new source domain features and target domain features to obtain intermediate domain features.Then the new source features,intermediate domain features and target features are used for adversarial training in discriminator.We conduct experiments on MNIST,USPS,and SVHN datasets.The experiments verify the effectiveness of the proposed method,and also analyze other indicators such as the sensitivity of the method.And the comparison experiments with other methods show the superiority of NMADA.Secondly,according to the characteristics of medical recognition tasks,we adjust the adversarial framework NMADA to carry out computer medical aided diagnosis.The method first preprocesses the medical images,initializes the model with pretrain model parameters using source domain data,then comprehensively considers the low-level features and high-level features,seeks more intermediate domain features and common features,and matches effectively the data distribution.Medical classification experiments are carried out on medical datasets composed of viral pneumonia and COVID-19 medical images to verify the effectiveness of method.The performance of method under different noise scenarios is evaluated,which proves that it still has good performance in the complex medical environment.