Visual Emotion Analysis of Female Images Using Convolutional Neural Networks

Facial emotion recognition plays a crucial part in the realm of affective computing and human–computer interaction and is pivotal with respect to the understanding of an individual’s emotional state through their facial expressions. Existing approaches to emotion recognition are challenged by gender-based facial expression discrepancies, particularly with more subtle expressions in females. We present a novel framework for visual emotion analysis of female facial expressions in this paper. The proposed system consists of descriptions for the architecture of a multi-stream CNN, the mechanism of double attention for fusion of emotion-specific regions, and Supervised Contrastive Learning (SupCon) for a better feature representation. The multi-stream CNN network handles the original, augmented, and ROI-only images to extract global and local emotion-related features. The dual-attention mechanism maintains the focus on the most important regions and channels of the human face, in order to maintain and strengthen the robustness and interpretability of the models. SupCon will enhance inter-class isolability and intra-class compactness significantly; this brings a better emotion classification. By using the classification model configuration, it was possible to achieve 99.30% classification accuracy with 98.13% precision, 98.70% recall and Cohen’s Kappa of 0.9916 — very close to the actual labels provided. Also, FNR was insignificant at 0.0010% and FPR was 1.30, which means that the system is highly specific and the strongest in terms of analyzing facial emotions in gender.