Research On Unsupervised Domain Adaptation Algorithm Based On Adversarial Learning

With the generation of large-scale data and the improvement of computer hardware performance,deep learning algorithms have achieved rich results in computer vision tasks such as image classification,semantic segmentation and object detection.However,there is often a lack of sufficient labeled data in many computer vision tasks.Affected by the environment and equipment,the training image data of the source domain and the test image data of the target domain usually do not meet the conditions of independent and same distribution.These factors impair the generalization performance of the source domain training model on the target task.Domain adaptation can extract similar knowledge from the source domain and the target domain,and apply the learned knowledge to the target task,thereby reducing the distribution difference between different domains.In recent years,a large number of adversarial domain adaptation studies have focused on how to use adversarial learning methods to align data distribution between domains,but there are still many problems : 1)How to extract category-level feature representations in the process of adversarial domain adaptation;2)Bi-classifier adversarial domain adaptation method only considers minimizing the prediction difference between the two classifiers on the target sample,and cannot guarantee the prediction accuracy of the target sample.In view of the above problems,two new unsupervised domain adaptation algorithms based on adversarial learning are proposed,and these two methods are summarized as follows:Aiming at the scenarios of unsupervised domain adaptation and single-adversarial domain adaptation,an adversarial domain adaptation model based on feature anomaly detection and pseudo-label regression is proposed,which solves the problem that the pseudo-label data of the target domain contains a lot of noise and the difficulty of representing category-level features in the process of adversarial domain adaptation.This method uses the target sample of the same class predicted by the classifier as the subdomain of the target domain,detects the target feature with anomalous distance from the class mean feature in each subdomain by the Gaussian uniform mixing model,outputs the posterior probability of the sample to measure the correctness of the pseudo-label in the subdomain,and limits the influence of the pseudo-label data of the target domain on the domain adaptation model in training as a pseudo-label multiplier.It is proposed that pseudo-label regression loss reduces the difference between the classifier prediction of the target sample and its high-confidence pseudo-label,so as to improve the class discrimination of the target sample.It is proposed to minimize the entropy loss learning feature to stay away from the decision boundary,and further improve the classification performance of the network.In addition,the target domain pseudo-labels with high confidence are fine-tuned as trainable data to the domain adaptation model.Aiming at the scenarios of unsupervised domain adaptation and bi-classifier adversarial domain adaptation,a bi-classifier adversarial domain adaptation model based on domain gradient difference and self-supervised learning is proposed,which solves the problem of how to ensure the prediction accuracy of the target sample when minimizing the prediction difference between the two classifiers.This method consists of two parts: model pre-training and bi-classifier adversarial training.The prediction consistency of two classifiers and weighted clustering on target domain samples is used to obtain pseudo-labels,and the self-supervised loss pre-training model is used to improve the prediction accuracy of the target samples by the two classifiers,so as to avoid negative transfer caused by classifier prediction errors in subsequent adversarial training.On the basis of minimizing the classifier difference,the difference between the gradient vector of the source classification loss and the target classification loss is used as a potential supervised information learning feature extractor,so as to generate a feature representation with domain invariance.In order to evaluate the effectiveness of the method proposed in thesis paper,the Office-31,Image-CLEF and Office-Home datasets were used for verification in the experiment.In addition,a variety of auxiliary experiments are designed to verify the effectiveness of the proposed models and the associated loss function.Experiments show that the proposed method can effectively reduce the difference of domain distribution.