判别与结构信息保持的无监督领域自适应方法  

Discrimination and structure preserved cross-domain subspace learning for unsupervised domain adaption

作　　者：陶洋[1] 杨娜 郭坦 TAO Yang;YANG Na;GUO Tan(College of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《西安电子科技大学学报》2022年第4期90-99,117,共11页Journal of Xidian University

基　　金：重庆市自然科学基金面上项目(cstc2020jcyj-msxmX0636);重庆市“成渝地区双城经济圈建设”科技创新项目(KJCXZD2020025);重庆市技术创新与应用发展专项面上项目(cstc2020jscx-msxmX0178)。

摘　　要：基于特征表示的领域自适应是研究的最广泛的迁移学习方法。但是,目前的方法在获得新的数据表征之后,没有考虑类内与内间的关系,并且在转换的过程中丢失了数据的几何结构信息。针对上述问题,提出一种判别与结构信息保持的无监督领域自适应方法。该方法在低秩子空间学习的框架下,从保留源域的判别信息和数据的结构信息两方面考虑,提高分类器在目标域样本的分类准确度。具体地,该方法寻找一个源域和目标域之间的不变子空间,通过低秩约束的跨领域样本重构学习以减少跨域分布差异,在迁移过程中,使用标签松弛矩阵最大化源域不同类别样本的类间距离和源域类间稀疏约束缩小来自同类别样本的距离,有效地保留了源域的判别信息。同时,利用自适应概率图结构保留样本的局部近邻关系,捕获数据底层的几何结构信息,增强子空间学习的鉴别力和鲁棒性。在3种不同的跨域图像基准数据集上实验验证了提出方法的有效性。实验结果表明,所提算法的分类性能优于目前存在的方法。One popular transfer learning method is domain adaptation based on feature representation.However,such a method fails to consider the within-class and between-class relations after obtaining the new data representation.In addition,the geometric structure information of the data is lost during the conversion process.To overcome this problem,a novel discrimination and structure preserved cross-domain subspace learning for the unsupervised domain adaption method is developed in this paper.Under the framework of low rank subspace learning,this method improves the classification accuracy of the classifier in the target domain by considering both the discrimination information of the source domain and the structural information of the data.Specifically,the method finds an invariant subspace between the source domain and the target domain,and uses cross domain sample reconstruction learning with low rank constraints to reduce the difference of cross domain distribution.In the process of migration,the label relaxation matrix is used to maximize the inter class distance of samples of different categories in the source domain and the sparse constraint between classes in the source domain to reduce the distance from samples of the same category and effectively retain the discrimination information of the source domain.At the same time,the adaptive probability graph structure is used to retain the local nearest neighbor relationship of samples,capture the geometric structure information at the bottom of the data,and enhance the discrimination and robustness of subspace learning.Experiments on three different cross-domain image data sets verify the effectiveness of the proposed method.Experimental results show that the classification performance of the proposed algorithm is better than that of the existing methods.

关 键 词：迁移学习 子空间学习 领域自适应 类间稀疏 图结构 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]