基于渐进式双重对齐的无源无监督领域自适应方法

Gradual dual alignment for source-free unsupervised domain adaptation

无源领域自适应的核心任务是利用无标签的目标域数据,将预训练好的源模型迁移到目标领域。基于深度聚类的方法需要在自监督学习过程中挖掘辅助信息来正则化特征分布对齐,而辅助信息中噪声常常误导该对齐过程;基于伪源域的对抗学习方法进行概率分布对齐,对所构建伪源域质量十分敏感。针对现有方法存在的不足,本文提出了一种基于渐进式双重对齐的无源无监督领域自适应方法,在进行深度聚类的同时,进行域对齐,缓解深度聚类中伪标签的噪声。首先,通过超近邻增强样本生成高质量伪源域,以克服源域不可见的问题;其次,利用对抗学习,实现两个域的概率分布初对齐;最后,引入深度特征相似,进一步强化对齐效果。在两个公开数据集上的实验结果表明了其有效性。

The principle task of source-free domain adaptation is to transfer a pre-trained source model from the source domain to the target domain by only using unlabeled target domain data.The deep clustering-based methods mine the auxiliary information from the self-supervised learning process to regulate the feature distribution alignment.However,the noise in the auxiliary information always misleads this alignment.Besides,the adversarial learning-based methods,conducting a probability distribution alignment,are sensitive to the constructed pseudo source domain.Aiming at the two shortcomings,this paper proposes a new gradual dual alignment(GDA)approach,which performs deep clustering where the errors in pseudo-labels are alleviated by a domain alignment.Specifically,this hybrid framework adopts the classical paradigm to perform deep clustering.As for the domain adaptation,a dual alignment including probability distribution and feature alignment is developed:First,a pseudo-source domain is constructed by hyper-nearest neighbor sample generation to overcome the problem of the invisible source domain;followed the adversarial learning is used to achieve initial alignment of the probability distributions of the two domains;At the end,deep feature alignment is introduced to further enhance the alignment effect.Experimental results on two public datasets demonstrate its effectiveness.