基于原型引导与双重对比学习的无源域自适应OA

Source-free domain adaptation(SFDA)aims to enable effective knowledge transfer of pre-trained models from the source domain to the target domain when the source samples are inaccessible.Most SFDA methods adopt self-training learning,but they are susceptible to noisy pseudo-labels,making it difficult for models to accurately capture in-trinsic features and leading to performance degradation.To address this issue,we propose a novel SFDA method named prototype guidance and dual-level contrastive learning(PDCL).Specifically,PDCL first extracts features from the source domain using the pre-trained source model to construct initial class-level prototype representations for the target domain.Subsequently,it aligns the features of source and target samples via a feature fusion strategy,thereby generat-ing pseudo-labels for the target domain.Secondly,it introduces a prototype orthogonality loss to enhance discriminabil-ity by maximizing the orthogonality between class prototypes.Finally,PDCL employs a dual-level contrastive learning strategy,where instance-level contrast is conducted in the decision space to pull near neighbors closer and push outliers away,while prototype-level contrast is conducted in the feature space to enhance structural consistency.Experimental results on Office-31,Office-Home,and DomainNet-126 datasets show that PDCL outperforms state-of-the-art methods,which validates its effectiveness.