<p>In Facial Expression Recognition (FER) tasks, the long-tail distribution and the presence of hard samples pose significant challenges to recognition accuracy. Existing methods often focus excessively on dominant expression classes, resulting in suboptimal performance for minor expression classes, particularly in complex, real-world environments. To address these issues, this paper proposes RIC-FER (Rebalancing Imbalanced Classes in Facial Expression Recognition), a method designed to tackle both long-tail distribution and hard sample learning, thereby enhancing FER performance on imbalanced datasets. Our method consists of two main components: a feature augmentation module based on class activation maps and a robust feature mining module. The former integrates distinctive features of minor classes with general features of dominant classes to enrich the feature distribution of minor classes and improve their representational capacity. The latter employs a dynamic scoring mechanism to enhance the model’s ability to learn from difficult samples. Additionally, a multi-view debiasing loss is introduced, which jointly considers sample classification difficulty and class imbalance. This loss function optimizes the model’s learning process across categories, effectively mitigating the negative impact of data imbalance. Experimental results demonstrate that RIC-FER substantially improves the recognition performance of FER models, particularly for minor expression classes and hard samples, and outperforms existing methods for imbalanced data handling.</p>