Semantic Reconstruction Coupled Reliable Pseudo-Label Learning for Unsupervised Domain Adaptation
Mengtao YANG , Yueying LIU , Gaohao YI , Hao ZHOU , Yazhou YANG , Tingjin LUO
Unsupervised domain adaptation aims to train a high-performing classifier for the unlabeled target domain by exploiting fully labeled source data, despite the distribution discrepancy between domains. Existing methods typically employ pseudo-labeling to capture target semantic information, yet overlook how unreliable labels constrain performance. Besides, traditional approaches using all source samples for target reconstruction tend to blur category boundaries, degrading its learning discriminability. To address these challenges, we propose the Semantic Reconstruction coupled Reliable Pseudo-label Learning method named as SRRPL in this paper. Specifically, a ”two-step pruning” pseudo-label strategy is adopted to enhance reliability through nearest-neighbor screening and probability pruning, by enabling finer-grained characterization of target samples. Moreover, SRRPL integrates a class-level semantic reconstruction matrix that employs intra-class source samples as bases to reconstruct data from both source and target domains, thereby enhancing class discriminability. Besides, an efficient alternating iterative optimization algorithm is designed to solve our formulated non-convex objective function. Finally, extensive experimental results on multiple public datasets demonstrate our method outperforms other state-of-the-art approaches.
Semantic Reconstruction / Pseudo-Label Learning / Subspace Learning / Unsupervised Domain Adaptation
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(4914KB)
