VLCE: Unified Vision-Language Collaborative Enhancement for Facial Expression Recognition

Facial Expression Recognition (FER) is a challenge task due to the inherent variability in facial expressions and the complex contextual factors in real-world data. Traditional models often focus solely on visual features, limiting adaptability in complex scenarios. While recent methods incorporate semantic information through text-based descriptions, they generally isolate vision and language modalities, restricting effective cross-modal integration. In this paper, we present a simple and effective Vision-Language Collaborative Enhancement (VLCE) framework for FER, which integrates vision and language features through unified collaborative learning. Our dual-encoder architecture, consisting of a CLIP-based textual encoder and a visual encoder, is jointly trained to learn class-specific semantic templates and visual features. By aligning text-based embeddings that capture nuanced expressions with visual representations, the framework enables effective cross-modal feature fusion, enriching visual features with semantic context. The VLCE framework demonstrates competitive performance on multiple benchmark datasets, with extensive ablation studies and visualizations validating its robustness and interpretability, bridging the gap between visual and textual modalities in FER.