ACAForms: Learning Adaptive Context-Aware Feature for Facial Expression Recognition in Human-Robot Interaction

Facial expression recognition (FER) dedicates to achieve accurate facial expression classification and it is also a significant task in human-robot interaction. By understanding human emotions through facial expressions, robots can adapt their responses to fit the user’s mood or intent, enhancing the interaction experience. In this work, we propose a novel facial expression classification method based on Transformer for FER that combines CNN and Transformer architecture and design a unique feature relevance learning mechanism. Two modules are mainly contained in our model, namely, context-aware local feature learning (CALL) and Adaptive Feature Relationship Induction (AFRI). Specially, CALL is introduced to learn local features and use attention mechanisms to sense context, focus on local features of faces with internal symmetry and external agreeableness. AFRI, on the other hand, uses a Transformer-based structure to extract global features and capture long-range dependencies within the facial data, which is essential for understanding subtle expression variations in HRI scenarios. Comprehensive experiments demonstrate that our model can achieve state-of-the-art performance on RAF-DB and KDEF datasets with the accuracy of 89.44% and 95.00%.