电子学报2025，Vol.53Issue(1)：128-140,13.DOI:10.12263/DZXB.20240074
基于特征异常检测与伪标签回归的无监督对抗域适应
Feature Anomaly Detection and Pseudo-Label Regression for Adversarial Domain Adaptation
摘要
Abstract
In unsupervised domain adaptation tasks,the source and target domains usually do not satisfy the indepen-dent and identical distribution assumption.In order to generate the usable labels for the target domain,classical domain ad-aptation methods select the category with the highest prediction probability of the classifier as the pseudo-label of the target sample.Thus,the pseudo-label inevitably contains certain noise information,which may cause negative transfer to the do-main adaptation model.In addition,traditional adversarial domain adaptation methods usually consider the global distribu-tion between domains and ignore the category information of samples.How to extract discriminative category-level features in domain adaptation tasks is also an important problem.Therefore,an unsupervised adversarial domain adaptation method is proposed using feature anomaly detection and pseudo-label regression.The target samples of the same class predicted by the classifier are formed into the category subdomain within the target domain.The Gaussian uniform mixture model is used to detect the subdomain samples with abnormal distance from the class mean.The posterior probability of the samples is calculated and the correctness of the sample pseudo-labels in the subdomain is measured,which is used as a loss factor to limit the influence of pseudo-labels on the model in training.Meanwhile,the pseudo-label regression function is used to re-duce the difference between the predicted label and the high-confidence pseudo-label of the classifier.The category con-straint of the unlabeled target domain is adopted to improve the distinguishability of feature categories.Experimental results show that the average recognition accuracy of the proposed method on datasets Office-31,Image-CLEF,and Office-Home are 90.2%,89.6%,and 69.5%,respectively,which are all higher than the related popular algorithms.关键词
对抗域适应/特征检测/高斯均匀混合模型/伪标签回归/无监督学习/图像分类Key words
adversarial domain adaptation/feature detection/Gaussian uniform mixture model/pseudo-label regres-sion/unsupervised learning/image classification分类
信息技术与安全科学引用本文复制引用
潘杰,刘波,邹筱瑜..基于特征异常检测与伪标签回归的无监督对抗域适应[J].电子学报,2025,53(1):128-140,13.基金项目
国家自然科学基金(No.62176258,No.62273349,No.61806207) （No.62176258,No.62273349,No.61806207）
中央高校基本科研业务费专项资金项目(No.2021YCPY0111) National Natural Science Foundation of China(No.62176258,No.62273349,No.61806207) （No.2021YCPY0111）
Fundamental Research Funds for the Central Universities(No.2021YCPY0111) （No.2021YCPY0111）
