Study on emotion recognition bias in different regional groups

Human-machine communication can be enhanced by real-time recognition of spontaneous human emotional expressions, but performance is impacted by lighting, obfuscation, and cultural/regional variation. An emotion recognition model trained on one region (e.g., North America) may fail on another (e.g., East Asia). To address regional and cultural bias in facial emotion recognition, the authors propose a meta-model that fuses multiple emotional cues and features into a multi-cues emotion model (MCAM), integrating image features, action units, micro-expressions, and macro-expressions. Each cue represents distinct categories: fine-grained content-independent features, facial muscle movements, short-term expressions, and high-level expressions. Results show: (a) successful classification of regional facial expressions can be based on non-sympathetic features, (b) learning expressions of some regional groups can confound recognition of others unless relearned from scratch, and (c) certain facial cues/features preclude a perfect unbiased classifier. The authors posit that to learn certain regional expressions, other regional expressions must first be forgotten.

Martin Lukac, Gulnaz Zhambulova, Kamila Abdiyeva, Michael Lewis