Journal of Frontiers of Computer Science and Technology ›› 2023, Vol. 17 ›› Issue (5): 1168-1179.DOI: 10.3778/j.issn.1673-9418.2108111

• Artificial Intelligence·Pattern Recognition • Previous Articles     Next Articles

Multiple Adversarial Deep Domain Adaptation Model with Inter-class Difference Constraint

MA Na, WEN Tingxin, JIA Xu   

  1. 1. School of Business Administration, Liaoning Technical University, Huludao, Liaoning 125105, China
    2. School of Electronics and Information Engineering, Liaoning University of Technology, Jinzhou, Liaoning 121001, China
  • Online:2023-05-01 Published:2023-05-01

具有类间差异约束的多对抗深度域适应模型

马娜,温廷新,贾旭   

  1. 1. 辽宁工程技术大学 工商管理学院,辽宁 葫芦岛 125105
    2. 辽宁工业大学 电子与信息工程学院,辽宁 锦州 121001

Abstract: To align the samples in the target domain to the samples of the same class in the source domain accurately, and improve the feature distinguishability between different classes further while ensuring recognition accuracy, a domain adaptation model with inter-class difference constraint is proposed. Firstly, a deep neural network is used to perform supervised learning on source domain samples, and the inter-class difference constraint is imposed on the sample features in the source domain based on the proposed inter-class difference measurement function. Then, multi-adversarial domain discrimination network is adopted, in which a pseudo label solving method is proposed to assign the unlabeled samples to the reasonable domain discriminators for training. Finally, the best feature extractor and feature classifier can be obtained through minimizing the classification loss and maximizing the domain discrimination loss. Experimental results show that for the four datasets, the average recognition accuracy of the proposed model on the target domain can reach 0.860, and the average distance between the same classes, the average distance between different classes, the error recognition rate in the target domain can be reduced by 0.003, improved by 0.065, and reduced by 0.025, respectively, which is verified that the performance of proposed model is improved significantly.

Key words: transfer learning, deep domain adaptation, inter-class difference, multiple adversarial networks

摘要: 为实现目标域样本能够与源域中同类样本准确对齐,并在保证样本准确识别率的条件下进一步提高不同类别样本特征间的可区分性,提出了一种带有类间差异约束的域适应模型。首先,该模型采用深度卷积神经网络对源域样本进行了有监督学习,并在训练过程中基于提出的类间差异测量函数对源域样本特征加以类间差异性约束;其次,该模型采用了多对抗域鉴别网络结构,其中提出了一种目标域样本伪标签计算方法,从而将无标签的样本指定到合理的域鉴别网络进行训练;最后,通过最小化分类损失与最大化域鉴别损失,获得最优特征提取器与特征分类器。实验结果表明,对于4种数据集,提出的模型在目标域上平均识别准确率可以达到0.860,同类间的平均距离、不同类间的平均距离、目标域中样本错误识别率相对于改进前分别降低0.003,提升0.065,降低0.025,从而验证了提出模型的性能得到了明显提升。

关键词: 迁移学习, 深度域适应, 类间差异, 多对抗网络