Facial Action Unit-Based Valence-Arousal Estimation with Subjective Validation

Facial expressions are fundamental nonverbal cues for conveying human emotional states, and their automatic analysis has become an important research topic in affective engineering and human-computer interaction. The Facial Action Coding System (FACS) provides an objective and quantitative framework for describing facial expressions on the basis of facial muscle movements, known as action units (AUs). However, the extent to which emotional states estimated from AU information correspond to subjective human perceptions has not been sufficiently examined. In this study, we developed a regression model that estimates valence-arousal coordinates from AU intensities using Russell’s circumplex model of affect. To evaluate the validity of the estimated emotional states, a subjective evaluation experiment using the emotion circumplex was conducted, and the agreement between the model predictions and participants’ judgments was quantitatively analyzed. The results show that the proposed AU-based model achieves stable estimation performance, and the level of agreement varies across emotional directions. The method developed in this study presents a human- centered framework for examining the psychological validity of AU-based emotion estimation models.