聚类中心对齐的无监督域适应网络  

Centroid Alignment for Unsupervised Domain Adaptation

作　　者：陈辛怡 孙涵[1] CHEN Xin-yi;SUN Han(College of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211106

出　　处：《小型微型计算机系统》2022年第4期822-827,共6页Journal of Chinese Computer Systems

基　　金：国防科技创新特区项目资助。

摘　　要：深度学习在图像分类上的准确度很大程度上依赖于大量的标记数据,无监督域适应已经被证明是一种有效的方法去解决一个新的无标签域上的任务,其主要思想是利用有标签的数据集作为源域,通过减少源域和目标域之间的差异,将源域训练的预测模型应用于目标域.本文提出了聚类中心对齐的无监督域适应方法CADA,将语义对齐方法与传统对抗域适应相结合.CADA首先在对抗训练中对齐两个域的特征空间的边缘分布,再经过对源域特征的中心增强操作,以及集成分类器为目标域样本分配伪标签,最后将源域中心和伪标注后的目标域中心进行对齐,达到语义迁移的效果.本文在office-31以及数字数据集上进行了实验,并与多种域适应方法进行了对比,结果表明CADA可以有效提高域适应效果并且在不同的应用场景中表现优异.The accuracy of deep learning on image classification tasks relies on a large amount of labeled data,and unsupervised domain adaptation has been shown to be an effective way to solve the task on a new unlabeled domain.Domain adaptation focuses on utilizing a labeled dataset as the source domain,and applying the prediction model trained on the source domain to the target domain by reducing the difference between the two domains.In this paper,we present Centroid Alignment for Unsupervised Domain Adaptation(CADA),which combines semantic alignment with traditional adversarial domain adaptation.CADA first match the marginal distributions of the source and target domains by adversarial training.Then in order to achieve semantic transfer,we execute center augmentation for source domain to align the labeled source domain centers and target domain centers labeled by ensemble classifier.We evaluate the performance of the proposed method for cross-domain image classification tasks on office-31 and digits datasets,CADA is robust to various scenarios,performs well across significant domain gaps,and remarkably outperforms contemporary domain adaptation methods.

关 键 词：域适应 对抗学习 语义迁移 集成学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]