摘要
针对无源域数据的域自适应问题中源模型中的源知识得不到充分利用以及目标数据内在结构信息会被忽略的问题,提出了一种基于不确定性指导的加权混合训练的无源域数据的域自适应框架。首先在通道级对源模型预测的不确定性进行量化,将目标数据同时输入到源特征提取器和目标特征提取器,在通道级计算目标特征与源模型之间的不确定性距离,衡量源知识对目标模型的可迁移的不确定性,并利用它来指导对目标数据的适应过程。同时考虑到目标数据的内在结构信息,从源知识和目标数据两个方面考虑伪标签的置信度,以降低噪声伪标签的影响。最后对目标样本进行混合训练,并根据置信度得分对目标样本进行加权,以提高目标模型的鲁棒性。在Office-31,Office-Home和VisDA-C 3个域自适应基准数据集上分别取得了90.5%,72.8%,85.9%的分类效果,大量的对比实验及消融分析都证明了所提出的基于不确定性指导加权混合训练的框架具备良好的迁移能力。
Aiming at the problem that the knowledge in the source model cannot be fully utilized and the intrinsic structure information of the target data is ignored in the source-free domain adaptation,which results in the performance degradation of the target model,a network based on uncertainty-guided weighted hybrid training is proposed.Firstly,the uncertainty of source model prediction is quantified at the channel level,the target data is fed into the source feature extractor and the target feature extractor at the same time,to calculate the uncertainty distance between the target feature and the source model at the channel level,and the transferability uncertainty of source knowledge to the target model is measured,which is used to guide the process of adapting target data.Considering the intrinsic structure of target data,the joint confidence score of the pseudo-label is considered from both source knowledge and target data to reduce the influence of the noise label.Finally,the target samples are conduced hybrid training,and weighted according to the confidence score to improve the robustness of the target model.The classification results of 90.5%,72.8%,and 85.9%on three domain adaptive benchmark datasets of Office-31,Office-Home,and VisDA-C,show that the proposed framework for source-free domain adaptation based on uncertainty-guided weighted hybrid training has a good transfer ability.
作者
关新宇
孙涵
GUAN Xin-yu;SUN Han(School of Computer Science and Technology/Artificial Intelligence,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)
出处
《计算机技术与发展》
2023年第11期135-142,共8页
Computer Technology and Development
基金
中央高校基本科研业务费专项资金资助项目(NZ2019009)。
关键词
域自适应
无源域数据
不确定性
伪标签
混合训练
domain adaptation
source-free data
uncertainty
pseudo label
hybrid training
