Integrated approach for simultaneous estimation of emotional states and valence-arousal levels in facial images

Recognizing and interpreting human emotions poses a significant challenge in affective analysis. Facial images serve as a major input for such systems, as facial expressions are a primary means of understanding human emotions. In this work, we address the challenge of accurate facial emotion recognition by presenting a pioneering method for simultaneously estimating emotion and valence-arousal levels from facial images. Our approach leverages a deep learning architecture with multiple output branches to predict emotional states and valence-arousal dimensions concurrently. The model is trained on the AffectNet dataset, which is annotated with discrete emotion labels and continuous valence-arousal scores. To enhance robustness and versatility, we incorporate data augmentation techniques during training. Empirical results demonstrate the effectiveness of our method in both emotion classification and valence-arousal prediction tasks. By jointly optimizing for emotion recognition and affective dimension estimation, our approach provides a comprehensive solution for interpreting human affective states through facial expressions, with promising applications in affective computing.