BFFN: A novel balanced feature fusion network for fair facial expression recognition

Abstract

Facial expression recognition (FER) technology has become increasingly mature and applicable in recent years. However, it still suffers from the bias of expression class, which can lead to unfair decisions for certain expression classes in applications. This study aims to mitigate expression class bias through both pre-processing and in-processing approaches. First, we analyze the output of existing models and demonstrate the existence of obvious class bias, particularly for underrepresented expressions. Second, we develop a class-balanced dataset constructed through data generation, mitigating unfairness at the data level. Then, we propose the Balanced Feature Fusion Network (BFFN), a class fairness-enhancing network. The BFFN mitigates the class bias by adding facial action units (AU) to enrich expression-related features and allocating weights in the AU feature fusion process to improve the extraction ability of underrepresented expression features. Finally, extensive experiments on datasets (RAF-DB and AffectNet) provide evidence that our BFFN outperforms existing FER models, improving the fairness by at least 16%.