Dual-Level Redundancy Elimination for Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) methods for image classification have rapidly advanced in recent years. This advancement is attributed to their ability to address performance degradation caused by distributional differences between source and target domain images. However, previous methods often focus solely on sample-level relationships, neglecting the problem of redundancy within samples due to the same feature information being expressed repeatedly. This oversight can smooth critical discriminative features, ultimately reducing the model’s accuracy in image classification. To address this problem, we propose Dual-Level Redundancy Elimination for Unsupervised Domain Adaptation (DLRE), which aims to enhance UDA by reducing redundancy at both the sample feature and cluster levels. Specifically, we use feature mutual information to evaluate and reduce the feature-level redundancy of different dimensions, thereby ensuring the amount of discriminative information contained in the feature vector, and maximize the dimensional mutual information between pairs of positive samples to obtain a more unbiased feature representation. In addition, we propose a new clustering framework based on dictionary storage and mutual information weighting to reduce cluster-level redundancy and help improve the performance of classification tasks. We conduct extensive experiments on five widely used UDA vision benchmarks, and the results show that DLRE has good adaptability and effectiveness, significantly outperforms current domain adaptation methods.