Edge-Optimized Facial Emotion Recognition: A High-Performance Hybrid Mobilenetv2-Vit Model

Computer vision applications span various fields including healthcare, security, autonomous vehicles, and augmented reality, enabling machines to interpret and analyze visual data. Facial Emotion Recognition (FER) is a subclass of healthcare applications that leverages computer vision to analyze and interpret human emotions from facial expressions. Facial emotion recognition also plays a vital role in human-computer interaction, with applications in security, and affective computing. This study suggests a deep learning (DL) based hybrid model integrating MobileNetV2 for efficient feature extraction and a Vision Transformer (ViT) for capturing global facial dependencies. The dataset obtained from Kaggle is used for training, which is then preprocessed and augmented. The trained model is deployed on a smartphone as an edge device, enabling real-time emotion recognition with improved privacy, low latency, and minimal computational overhead. During testing, facial images captured by the smartphone are preprocessed using the Haar Cascade algorithm before being fed into the model for classification. Performance evaluation using accuracy, recall, precision and F1-score demonstrates a high classification accuracy of 98.51%, confirming the model’s effectiveness. The proposed approach enhances on-device FER capabilities, making it a promising solution for emotion-aware applications in mobile healthcare and intelligent human-computer interactions