TransMCR: Learning Muscle Contraction Relationship for Facial Expression Recognition with Texture Deformation Representation

Facial expression recognition (FER) plays a vital role in areas such as human–robot interaction, security monitoring, and robot vision. However, FER encounters several challenges, including occlusions, lighting conditions and arbitrary face orientations. To tackle these challenges, we identify two cues from face images, namely muscle contraction relationships and texture deformation relationships. On the basis of the cues, three key anatomical insights from facial images are revealed: (i) facial muscle contraction relationship, (ii) regional coding of muscle interactions, and (iii) skin texture deformation relationship. To leverage three key insights above, a novel relationship-driven FER method is proposed based on Transformer architecture, in which muscle contraction and texture deformation relationships can be learned. Specifically, the Contraction Relationship Mining (CRM) is introduced to explore muscle contraction relationships through visual and contraction tokens. Then the Contraction Orientation Refinement (COR) scheme is developed to refine the relationship between facial muscles. Finally, the Texture Deformation Representation (TDR) module captures fine-grained superficial skin textures. Experiments on the RAF-DB, KDEF and FERPlus datasets show significant improvements of our proposed TransMCR framework over state-of-the-art FER methods. The source Python code is available upon request.

© 2026 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission.