Deep Hybrid Meta-Heuristic Model for Real-Time Facial Emotion Analysis

Facial expression recognition is important for human-computer interaction, as well as check in places such as healthcare and surveilling, yet real time recognition continues to be difficult due to lighting, pose, and intensity variation of expressions. This study proposes a Deep Hybrid Meta-Heuristic Model that incorporates Convolutional Neural Networks (CNN) and Vision Transformers (ViT) for multi-scale deep feature extraction, along with an Ant-Lion Optimization (ALO) algorithm to tune the hyperparameters and optimize classifier weight. We used the hybrid classification framework of SVM and Softmax layers in our proposed method to have multiple points of strength and increase generalization and robustness across the various benchmark datasets including the FER-2013 and RAF-DB datasets. The results indicated that the proposed method obtained 94.8% accuracy and 0.93 F1-score, above and beyond the conventional CNN, Transformer only, hybrid non-optimized, and baseline meta-heuristic models, while inference time was estimated at 28 ms per frame making the proposed method optimal for real time applications. Overall, our model along with these findings suggest that it is a solid, scalable, and efficient solution towards emotion-driven intelligent systems used within real-world varying environments.