Enhancing Robustness in Long-Tailed Image Classification with Angular Approaches and Balanced Learning

Abstract When the data distribution in a dataset is highly imbalanced or long-tailed, this can lead to biased deep network models that are ineffective at handling samples from the tail classes. The issue arises because the features of the samples in the head classes are dominant during the learning process. Supervised contrastive learning (SCL) and transfer of angular information (TAI) from head classes to tail classes are two methods proposed in recent literature. It was shown that, for a well-balanced dataset, SCL effectively improves class separation by pulling together samples from the same class while pushing away samples from different classes. However, SCL can remain biased towards the head classes since it does not directly address the effects of class imbalance during learning. TAI addresses challenges posed by long-tailed image classification but fails to achieve both intra-class compactness and inter-class separability. To address these limitations and exploit the strengths of both of these approaches, we propose a unique hybrid method that seamlessly integrates SCL and angular variance to mitigate the adverse effects of long-tailed data on deep learning models for image classification. The proposed method is named Supervised Angular Contrastive Loss with Balancing Classifier. Moreover, this paper finds that adding a suitably designed balancing technique can enhance results even further. Experiments on long-tailed datasets with different class imbalance ratios demonstrate that the proposed method significantly outperforms all existing baseline approaches.