CFE-FER: channel-guided feature enhancement for robust facial expression recognition

Accurate Facial Expression Recognition (FER) is crucial in sensitive applications such as mental health monitoring, driver surveillance, and human-computer interaction. Achieving robust FER in real-world applications requires lightweight models with strong inter-class separability and intra-class compactness, which depend on effectively capturing fine-grained facial expression features. Current state-of-the-art FER models often fail to generalize well due to the nuanced and complex nature of facial expressions, which are further affected by diverse real-world conditions such as partial occlusions, varying illumination levels, and demographic differences like race. To address these challenges, we propose CFE-FER, Channel- Guided Feature Enhancement for Robust Facial Expression Representation. The CFE-FER framework integrates three key modules: (1) Spatial Resolution Enhancement (SRE) to enrich spatial features and suppress redundant channel information, (2) Channel Attention (CA) to emphasize semantically meaningful channels, and (3) De-biasing Convolutional Refinement (DCR) to reduce identity-related and expression-independent biases. We extensively evaluated the proposed approach on three benchmark datasets: FERPlus, RAF-DB, and AffectNet. The results consistently demonstrate that our method outperforms state-of-the-art approaches, achieving statistically significant improvements with accuracies of 93.89% on FERPlus, 93.63% on RAF-DB, and 79.66% and 65.74% on the 7-class and 8-class settings of AffectNet, respectively.