ATR-Net: Attention-based temporal-refinement network for efficient facial emotion recognition in human–robot interaction

Facial Emotion Recognition (FER) enables human–robot interaction by allowing robots to interpret human emotions effectively. Traditional FER models achieve high accuracy but are often computationally intensive, limiting real-time application on resource-constrained devices. These models also face challenges in capturing subtle emotional expressions and addressing variations in facial poses. This study proposes a lightweight FER model based on EfficientNet-B0, balancing accuracy and efficiency for real-time deployment on embedded robotic systems. The proposed architecture integrates an Attention Augmented Convolution (AAC) layer with EfficientNet-B0 to enhance the model’s focus on subtle emotional cues, enabling robust performance in complex environments. Additionally, a Pyramid Channel-Gated Attention with a Temporal Refinement Block is introduced to capture spatial and channel dependencies, ensuring adaptability and efficiency on resource-limited devices. The model achieves accuracies of 74.22% on FER-2013, 99.14% on CK＋, and 67.36% on AffectNet-7. These results demonstrate its efficiency and robustness for facial emotion recognition in human–robot interaction.