Computer Science
Jyun-Yu Jhang; Cheng-Jian Lin; Xue-Qian Lin; Han Cheng
Articles in Press, Accepted Manuscript, Available Online from 01 October 2026
Abstract
As deep learning applications continue to expand, the performance of network models becomes increasingly critical. To enhance model accuracy, training often relies on a large amount of labeled data. However, manually labeling data is both labor-intensive and expensive. In recent years, Domain ...
Read More
As deep learning applications continue to expand, the performance of network models becomes increasingly critical. To enhance model accuracy, training often relies on a large amount of labeled data. However, manually labeling data is both labor-intensive and expensive. In recent years, Domain Adaptation (DA) techniques have gained significant attention as they enable models to be trained with limited labeled data while maintaining performance on unlabeled data. In particular, Unsupervised Domain Adaptation (UDA) faces a major challenge due to differences in data distributions. Addressing the problem of domain shift and identifying domain-invariant features with limited information remains a crucial task. This paper proposes a fuzzy-theory-based unsupervised domain adaptation approach called the Deep Fuzzy- Similarity-based Local Domain Adaptation Network (DFS-LDAN). The proposed method aims to mitigate domain shift and reduce annotation costs. DFS-LDAN integrates fuzzy similarity and Maximum Mean Discrepancy (MMD), allowing the model to retain more soft information during learning. This enhances its flexibility in handling uncertainties in target domain data. Additionally, this study introduces a Local Domain Adaptation strategy that aligns distributions at the class level. It employs Fuzzy C-Means (FCM) clustering to adaptively adjust feature weights based on membership degrees, further improving the model’s generalization capability. The effectiveness of DFS-LDAN is evaluated on three public datasets: Office-31, ImageCLEF-DA, and Office-Home. Comparative experiments with various domain adaptation methods demonstrate the superiority of DFS-LDAN. The results show that the proposed method, across the best-performing variant, achieves average accuracies of 89.1% (via Product-Sum), 90.4% (via Product-Sum), and 68.4% (via Min-Max), respectively.