基于特征和类别对齐的领域适应算法  

Domain adaptation based on feature-level and class-level alignment

作　　者：赵小强[1,2,3] 蒋红梅[1,2,3] ZHAO Xiao-qiang;JIANG Hong-mei(College of Electrical and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China;Gansu Key Laboratory of Advanced Control for Industrial Processes,Lanzhou University of Technology,Lanzhou 730050,China;National Experimental Teaching Center of Electrical and Control Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学电气工程与信息工程学院,兰州730050 [2]兰州理工大学甘肃省工业过程先进控制重点实验室,兰州730050 [3]兰州理工大学国家级电气与控制工程实验教学中心,兰州730050

出　　处：《控制与决策》2022年第5期1203-1210,共8页Control and Decision

基　　金：国家自然科学基金项目(61763029,61763027);国防基础科研项目(JCKY2018427C002);甘肃省高等学校产业支撑引导项目(2019C-05);甘肃省工业过程先进控制重点实验室开放基金项目(2019KFJJ01)。

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,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进行改进,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和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,为了能更深层次地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.所提出的算法在数据集Office-31上平均准确率为88.6%,在数据集Office-Home上平均准确率为67.7%,并与其他算法相比,验证了所提算法具备良好的泛化能力,可以实现较高的分类性能.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 maximum mean discrepancy(MMD)measurement criteria,a new improved maximizes the intra-domain density(MID)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.Then,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 67.7%for the dataset Office-Home.Compared with other algorithms,the proposed algorithm has better generalization ability and higher classification performance.

关 键 词：领域适应 联合判别网络 残差校正块 MID度量函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]