Research On Unsupervised Domain Adaptation Base On Generative Adversarial Network

The deep feedforward architecture brings impressive advanced technology to various tasks in computer vision and other fields.But these performance leaps only occur when a large amount of labeled training data is available.Due to the phenomenon of data set deviation or domain shift,the recognition model trained with these representations on a large data set cannot be easily extended to new data sets and tasks.Although these networks can be further fine-tuned on task-specific data sets,it is often very difficult and expensive to obtain sufficient labeled data to properly fine-tune the large number of parameters used by deep multilayer networks.In response to the above problems,this paper combines deep generation adversarial networks to study the unsupervised domain adaptation method,so that the deep model trained on the labeled source domain data can be applied to the unlabeled target domain with maximum accuracy,compared with other classic methods,And test and verify on different data sets.The main research contents of this article include:First,based on the unsupervised domain adaptation of the deep network,an adaptive layer is added on the basis of the Alex Net network,and the domain confusion loss and regularization term with the largest average difference are introduced to learn the semantically invariant domain representation.Train the deep network by minimizing the classification difference and maximizing the joint loss of domain loss,and use the domain confusion metric to select the optimal size of the adaptation layer and fine-tune the representation.Three transfer tasks were evaluated on the domain adaptation benchmark Office data,and how MMD regularization affected learning.Then,based on the idea of generating an adversarial network,the adversarial loss learning representation is introduced on the basis of the Alex Net network.The adversarial network can generate very realistic images through adversarial training,and the adversarial loss can more intelligently measure the distance between the domains without weight sharing.At the same time,the auxiliary classification task is introduced to maximize the training samples and enhance the generalization performance of the learned features.Experiments were carried out on digital data sets and remote sensing data sets respectively,and compared with the accuracy of other classic methods,and the confusion matrix was given on the remote sensing data sets.Finally,considering the task-specific decision boundaries between classes,the distribution of source and target domains is aligned by using task-specific decision boundaries.Use two classifiers to obtain features from the generator,correctly classify the source samples,and train them at the same time to detect target domain samples that are not included in the source domain support.Measure the difference between the two classifiers and train the generator to minimize the difference and avoid the generator generating target features outside the support range of the source domain.In order to achieve better performance of the network,batch normalization is used,so that the value of each layer is passed within the effective range.The experiment is adapted on the digital data set and the synthesized image and the real image,and the difference loss of each adaptation and the accuracy change of the 2 classifier are obtained,and the feature visualization is obtained.