HUMAN MACRO-EMOTION RECOGNITION FROM FACIAL IMAGES USING A DEEP CONVOLUTIONAL NEURAL NETWORK

Abstract

Facial emotion recognition underpins a growing range of human-computer interaction systems, yet reliable classification of expressions from photographic images remains difficult because of variation in expressive intensity, illumination, and imaging conditions. This paper reports a comparative study of two deep learning approaches to seven-class macro-emotion recognition on the Japanese Female Facial Expression (JAFFE) corpus: a transfer-learned MobileNetV3 backbone and a Deep Convolutional Neural Network (Deep CNN) designed and trained from scratch for this domain. Both models were trained under an identical protocol so that the comparison isolates architecture, and both were evaluated using precision, recall, F1-score, and overall accuracy, with per-class figures reported alongside aggregate ones. MobileNetV3 attained an overall accuracy of 94%, with markedly uneven per-class behaviour and its weakest performance on fear, sadness, and disgust. The proposed Deep CNN attained an overall accuracy of 99.56% with every class exceeding 0.98 on all three measures, and a component-wise ablation attributed the largest share of this margin to data augmentation, followed by convolutional depth, batch normalisation, and dropout. A computational analysis shows the compact backbone remains substantially cheaper at inference, so the choice between the two is best framed as a deployment decision rather than a ranking. The results indicate that a domain-specific network trained at higher input resolution recovers the fine-grained local evidence on which the clinically and interactionally important expression categories depend.