Mitigating Expression Class Bias with Class-Incremental Learning in Facial Expression Recognition
Yiqin Luo, Yinghui Li, Tianlong Gu, Liang Chang · 2025
As Facial Expression Recognition (FER) systems become increasingly integrated into daily life, ensuring fairness in decision-making is crucial, rather than focusing solely on improving recognition performance. FER systems need to find a trade-off between performance and fairness in order to avoid bias to subgroups while ensuring utility. To address this issue, various methods have been proposed at the data and algorithmic levels. However, these studies mainly focus on the unfairness of demographic attributes, and the bias of expression class remains largely unexplored. In this study, we attempt a novel strategy for applying class-incremental learning (class-IL) to mitigate expression class bias. Furthermore, we propose a fair model based on the characteristic of class-IL, called Incremental Expression Balance Network (IEBN). IEBN mitigates this class bias by adjusting the data distribution of different expressions and extracting critical features with an attention mechanism. Finally, experiments on the RAF-DB and AffectNet datasets show that IEBN outperforms other methods in terms of fairness and trade-offs between fairness and performance.