Multi-Label Facial Emotion Classification by Efficient Network from Korean Drama Video Clips

This paper presents an EfficientNetV2-based framework for multi-label facial emotion recognition tailored to Korean facial expressions. Unlike traditional single-label models trained on Western datasets, our method embraces the complexity of co-occurring emotions by modeling them as concurrent affective states. Utilizing a dataset of 38,817 annotated images from Korean dramas, we fine-tune a deep network with gradual decoding layers and balanced data augmentation for multi-label facial emotion classification. Experimental results show state-of-the-art performance under various thresholds, achieving 95.80% accuracy and 97.27% F1-score at threshold = 0.5, significantly surpassing baseline models. The proposed framework offers a robust and culturally aware solution for affect-aware AI systems operating in socially rich environments.