| 基于不对称对抗和自监督增强的无监督域自适应算法 |
| 点此下载全文 |
| 引用本文：尤雨婷,谭金菁,高瑞玲,毛文逸,吕金龙,侯英琦,谭文安.基于不对称对抗和自监督增强的无监督域自适应算法[J].上海第二工业大学(中文版),2025,42(3):298-306 |
| 摘要点击次数: 507 |
| 全文下载次数: 73 |
|
| 基金项目:国家自然科学基金项目(61672022, U1904186), 上海市研究生教育学会研究课题(ShsgeG202207) 资助 |
|
| 中文摘要:针对现有无监督域自适应(unsupervised domain adaptation, UDA) 算法在处理复杂领域分布差异时存在的特征对齐细粒度不足、目标域潜在特征挖掘能力有限等问题, 提出了一种基于不对称对抗和自监督增强的无监督域自适应(UDA based on asymmetric adversarial learning and self-supervised enhancement, UDA-2A2S) 算法。该算法通过引入全局与局部域判别器协同作用机制用于捕捉细粒度的领域特征差异, 并结合渐进对齐策略动态调整特征对齐强度, 以更好地适用于复杂任务场景。此外, 设计了一种基于旋转预测的自监督学习机制, 从目标域无标注数据中挖掘潜在的图像实例结构信息, 进一步提升模型的特征学习能力和跨域迁移能力。实验结果表明, 在Office-Home 和DomainNet 基准数据集上, UDA-2A2S 方法平均准确率分别达到72.4% 和46.2%, 显著优于现有主流UDA 方法, 验证了所提算法的优越性, 为跨域迁移学习提供了新的解决方案。 |
| 中文关键词:无监督领域自适应 不对称对抗机制 自监督增强 迁移学习 |
| |
| An Unsupervised Domain Adaptive Algorithm with Asymmetric Adversarial and Self-Supervised Enhancement |
|
|
| Abstract:To address the limitations of existing unsupervised domain adaptation (UDA) algorithm in handling complex domain distribution differences, including the insufficient fine-grained feature alignment and limited ability to extract potential features in the target domain, the study proposed an UDA based on asymmetric adversarial learning and self-supervised enhancement (UDA-2A2S) algorithm.
The proposed UDA-2A2S is used to capture fine-grained domain feature differences by introducing a synergistic mechanism of global and local domain discriminators, and dynamically adjusts the feature alignment strength in combination with a progressive alignment strategy to better apply to complex task scenarios. In addition, a self-supervised learning mechanism based on rotation prediction
was designed to mine potential image instance structure information from unlabeled data in the target domain to further enhance the feature learning capability and cross-domain migration ability of the model. The experimental results show that the UDA-2A2S method has an average accuracy of 72.4% and 46.2% on the Office-Home and DomainNet benchmark datasets, respectively, which is
significantly better than the existing mainstream UDA methods, demonstrating the superiority of the algorithm and providing a new solution for cross domain transfer learning. |
| keywords:unsupervised domain adaptation asymmetric adversarial mechanisms self-supervised enhancements transfer learning |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|