Luo Weibin, Chen Mingye, Gao Jian, Zhu Yanping, Wang Fang, Zhu Chenyang
School of Computer Science and Artificial Intelligence, Changzhou University, Changzhou, 213000, China.
Department of Computer Science, Brunel University London, London, UB8 3PH, UK.
Sci Rep. 2025 Jul 1;15(1):20452. doi: 10.1038/s41598-025-05331-3.
Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain with differing data distributions. However, it remains difficult due to noisy pseudo-labels in the target domain, inadequate modeling of local geometric structure, and reliance on a single input view that limits representational diversity in challenging tasks. We propose a framework named Multi-view Affinity-based Projection Alignment (MAPA) that uses a teacher-student network and multi-view augmentation to stabilize pseudo-labels and enhance feature diversity. MAPA transforms each sample into multiple augmented views, constructs a unified affinity matrix that combines semantic cues from pseudo-labels with feature-based distances, and then learns a locality-preserving projection to align source and target data in a shared low-dimensional space. An iterative strategy refines pseudo-labels by discarding low-confidence samples, thereby raising label quality and strengthening supervision for the target domain. MAPA also employs a consistency-weighted fusion mechanism to merge predictions from multiple views, improving stability under domain shift. Finally, MAPA leverages class-centric and cluster-level relationships in the projected space to further refine label assignments, enhancing the overall adaptation process. Experimental results on Office-Home, ImageCLEF, and VisDA-2017 show that MAPA surpasses recent state-of-the-art methods, and it maintains robust performance across backbones including ResNet-50, ResNet-101, and Vision Transformer (ViT).
无监督域适应（UDA）旨在将知识从有标签的源域转移到数据分布不同的无标签目标域。然而，由于目标域中存在噪声伪标签、局部几何结构建模不足以及依赖单一输入视图限制了挑战性任务中的表示多样性，这一过程仍然困难重重。我们提出了一个名为基于多视图亲和性的投影对齐（MAPA）的框架，该框架使用师生网络和多视图增强来稳定伪标签并增强特征多样性。MAPA将每个样本转换为多个增强视图，构建一个统一的亲和矩阵，将来自伪标签的语义线索与基于特征的距离相结合，然后学习一个保持局部性的投影，以在共享的低维空间中对齐源数据和目标数据。一种迭代策略通过丢弃低置信度样本细化伪标签，从而提高标签质量并加强对目标域的监督。MAPA还采用了一种一致性加权融合机制来合并来自多个视图的预测，提高域转移下的稳定性。最后，MAPA利用投影空间中以类为中心和聚类级别的关系进一步细化标签分配，增强整体适应过程。在Office-Home数据集、ImageCLEF数据集和VisDA-2017数据集上的实验结果表明，MAPA优于最近的先进方法，并且在包括ResNet-50、ResNet-101和视觉Transformer（ViT）在内的各种骨干网络上都保持了稳健的性能。
