Research And Application Of Unsupervised Domain Adaptation Algorithm Based On Adversarial Training

Deep neural network has achieved great success in image classification,object recognition and other tasks.However,the performance of the model will be greatly reduced when the data distribution of the training set(source domain)and the test set(target domain)is different.Domain adaptation has important practical significance in solving this problem.This paper focuses on learn domain invariant features without using the target domain label to improve the classification accuracy of the target domain.Existing domain adaptation methods usually ignore the classification information of the target sample.In order to "cheat" the discriminator,the generator often generates ambiguous features near the decision boundary,which affects the classification performance of the target domain.This paper proposes a domain adaptation model based on adversarial training.The classifier is used as a discriminator.While promoting feature alignment,the relationship between the target feature and the decision boundary is fully considered.In addition,the target domain classification information is used for model optimization.The classifier detects the classification boundary,and the generator generates discriminative features far away from the boundary.This article mainly completes the following tasks:1.From the perspective of unsupervised domain adaptation,this paper analyzes the research hotspots in recent years,and analyzes the optimization direction of unsupervised domain adaptation algorithm from the perspective of theory.The advantages and disadvantages of the existing unsupervised adversarial models are introduced in detail.Based on the above research,several domain adaptation methods based on adversarial training are proposed to extract domain invariant features to improve the robustness of the model.2.An unsupervised domain adaptation method based on adversarial entropy is proposed.This proposed method fully considers the discriminative information of the target domain.We achieve the following three goals:(1)Realizing the separability of the target sample in the feature space through maximum and minimum conditional entropy of the target samples by the adversarial training between a feature generator and a multi-class discriminator.(2)Obtaining the classification diversity by maximizing the marginal entropy of the target samples.(3)The label information of the target domain sample is generated to encourage the generator to output more discriminative features.Experiments show that our method has achieved good performance in different cross-domain experiments.3.An unsupervised domain adaptation method based on adversarial ensemble classifiers is proposed.This proposed method achieves the following three goals:(1)Learning the discriminative feature information and classification boundary of the source domain samples.(2)Exploring the classification boundary by maximizing the discrepancy of classifiers,and then minimizing the discrepancy of classifiers through the generator so as to achieve the alignment of the conditional distributions of the source domain and the target domain,which increases the separability between different classes.(3)The generator maximizes the mutual information between the input and output to increase the classification certainty and diversity of target domain samples.The rationality of the algorithm is proved theoretically.The experimental results on different tasks such as number recognition and target recognition show the effectiveness of our method.4.An image feature extraction method based on adversarial regularization is proposed.This proposed method achieves the following two goals:(1)The classifiers maximize the classification difference to improve the sensitivity,and the generator minimizes the difference to produce more robust features.(2)Exploring the influence of several difference measurement methods on the experiment.The experimental results on different datasets show that maximizing classifiers' discrepancy method is not only suitable for cross domain image classification,but also has good application in image feature extraction.