ER-EDF: An Enhanced Real-Time Emotion Detection Framework for Edge Devices in Emotional Robotics

Facial expression recognition (FER) is a pivotal task in artificial intelligence and affective computing, with broad applications in human-computer interaction, psychological analysis, and emotional robotics. However, FER models deployed on edge devices encounter critical challenges, including unbalanced datasets, limited generalization capabilities, and the trade-off between real-time efficiency and accuracy. To address these issues, this paper proposes the enhanced real-time emotion detection framework (ER-EDF), designed specifically for edge-device deployment in family companion robots. ER-EDF integrates three key innovations: First, Data Augmentation via Diffusion Models: Utilizing the Flux diffusion model [1], ER-EDF generates synthetic facial expression data with diverse attributes, such as age, gender, and scene-specific variations, effectively mitigating dataset imbalances and enhancing model generalization. Second, Feature Extraction with Optimized Attention Mechanisms: The framework incorporates the CBAM [2] and the novel universal feature attention module (UFAM) to emphasize discriminative features while compensating for neglected general features, improving both robustness and precision. Third, discriminative general feature loss function (DGFL): This loss function reduces intra-class variance and maximizes inter-class separability, enabling superior recognition of difficult expressions. Comprehensive evaluations on public FER datasets, including RAF-DB [3], FERPlus [4], and AffectNet-7 [5], demonstrate the effectiveness of ER-EDF. The framework achieves state-of-the-art accuracy while maintaining lightweight architecture and low computational overhead, making it ideal for real-time applications. These results highlight the potential of ER-EDF to advance FER capabilities in embedded systems and emotional robotics.