Cycle-reconstructive subspace learning with class discriminability for unsupervised domain adaptation. (September 2022)

• Record Type: Journal Article
• Authors: Xu, Yayun; Yan, Hua
• Is Part Of: Pattern recognition. Volume 129(2022)
• Publication Date: 2022-09
• Subjects: Domain adaptation -- Subspace learning -- Transfer learning -- Knowledge transfer; 97R40 -- 68T10; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108700
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Present an effective subspace learning approach for cross-domain image classification of unlabeled target objects. Two reconstructive matrixes are used through an iterative strategy to cycle-reconstruct data matrixes and update the common subspace. Different constraints are imposed on different reconstruction matrixes to preserve different structure information of the original domains. Considering the limitations of subspace learning, class discriminative constraints are added to improve recognition accuracy. Sufficient experiment results show that our proposed traditional method outperforms state-of-the-art traditional methods and is comparable with advanced deep methods on five current domain adaptation datasets. Abstract: Unsupervised domain adaptation is used to effectively learn a classifier for data of the unlabeled target domain by utilizing the data of the source domain with sufficient labels but different distributions. In general, a transformation matrix is employed to acquire a common subspace where the distributions of the two domains are aligned, which is easy to lose lots of unique information of the two domains. To better preserve useful information during the transformation process, we propose a novel Cycle-Reconstructive Subspace Learning with Class Discriminability (CRSL) approach that uses two reconstructive matrixes through an iterative strategy to cycle-reconstruct data matrixes and update the common subspace. In this way, we learn the invariant features in the common subspace while better preserving global and local structures of the two original domains. Finally, we implement additional discriminative constraints such as intra-class aggregation and inter-class diffusion on the transformed features to ensure the class discriminability of data of the two domains. Extensive experiment results show that our conventional method outperforms state-of-the-art conventional methods and is comparable with advanced deep methods on four current domain adaptation datasets.