基于特征和类别对齐的领域适应算法
作者:
作者单位:

兰州理工大学电信学院

作者简介:

通讯作者:

中图分类号:

TP391.41

基金项目:

国家自然科学基金项目(Nos.61763029, 61873116)(面上项目,重点项目,重大项目),国防基础科研项目(Nos.JCKY2018427C002)


Domain Adaptation based on Feature-level and Class-level Alignment
Author:
Affiliation:

School of Telecommunications, Lanzhou University of Technology

Fund Project:

The National Natural Foundation of China under Grant Nos.61763029, 61873116, the National Defense Basic Research Project of China under Grant Nos.JCKY2018427C002

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    针对现有的基于对抗学习的领域适应算法未能充分挖掘样本的可转移特征导致泛化能力较差和分类精确度较低的问题,本文提出基于特征和类别对齐的领域适应(FCDA)算法。首先,针对最大均值差异(MMD)度量准则存在的不足,改进得到一种新的MID(Maximizes the Intra-domain Density)度量函数,分别度量具有相同标签的源域样本特征间的分布散度、相同标签的目标域样本特征间的分布散度,实现最大化域内同类样本的类密度,从而降低类的错分率;其次,为了能更深层次的学习目标样本的抽象的、可转移的特征,从而减小域间差异,在特征提取网络后加入残差校正块,深化基础网络,提高其特征的可迁移性;最后,将获取的特征经过联合判别网络,通过对抗损失函数同时实现在类级和域级的对齐。本文算法在数据集Office-31上平均准确率为88.6%,在数据集ImageCLEF-DA上平均准确率为89.7%,并与其它算法相比,验证了本文算法具备良好的泛化能力,可以实现较高的分类性能。

    Abstract:

    Aiming at the problems of existing domain adaptation algorithms based on adversarial learning that they cannot effectively learn transferable features and have poor generalization ability, a domain adaptation algorithm based on feature and category alignment (FCDA) is proposed in this paper. First of all, in view of the shortcomings of the MMD measurement criteria, a new improved MID (Maximizes the Intra-domain Density) measurement function is obtained, which measures the distribution divergence between the source domain sample features with the same label, and the distribution divergence between the target domain sample features with the same label, so as to maximize the class density of similar samples in the domain, thereby the class error rate is reduced. Secondly, in order to learn the abstract and transferable features of the target sample at a deeper level, and reduce the difference between domains, a residual correction block is added after the feature extraction network to deepen the basic network, and the transferability of its features is improved. Finally, the acquired features are passed through the joint discriminant network, and the alignments at the class-level and the domain-level are achieved with the adversarial loss function. The proposed algorithm has an average accuracy of 88.6% for the dataset Office-31 and an average accuracy of 89.7% for the dataset ImageCLEF-DA. Compared with other algorithms, the proposed algorithm has better generalization ability and higher classification performance.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2020-09-07
  • 最后修改日期:2021-03-08
  • 录用日期:2021-03-16
  • 在线发布日期:
  • 出版日期: