Learning long-tailed representations via diversity-enhanced contrastive learning

A key challenge of long-tailed classification lies in the insufficient diversity of minority class samples and their learned representations. Existing long-tailed learning methods attempt to mitigate these issues by re-sampling or re-weighting strategies, but fail to enrich the diversity of minority classes. In this manuscript, we propose diversity-enhanced contrastive learning (DECL), a novel framework designed to improve representation learning and model performance on long-tailed classification tasks. Specifically, DECL consists of three key modules: a diversity-enhancing image augmentation (DIA) module, a feature extraction module and a diversity-aware contrastive learning (DCL) module. The DIA module focuses on increasing both the diversity and quantity of minority class samples, enriching the sample space. Then, the feature extraction module extracts diverse features from the augmented samples. Finally, the DCL module ensures clearer inter-class boundaries by aligning features within the same class and maximizing separation between different classes. By integrating these modules, the proposed method effectively mitigates the challenges associated with imbalanced datasets, enhancing feature diversity, and improving model generalization. Experimental results on four prevailing long-tailed benchmark datasets demonstrate that our method outperforms state-of-the-art competitors. The code will be made public at https://github.com/ZPlus233/DECL.