A Novel Transformer-Based Approach for Adult's Facial Emotion Recognition

Adult facial expression recognition (FER) is essential for human-computer interaction, mental health assessment, and social robotics applications because it improves user experiences and emotional well-being. This study presents a novel attention mechanism-based transformer approach designed to capture detailed patterns in facial features and dynamically focus on the most relevant regions for enhanced accuracy. Unlike conventional deep learning approaches, our method integrates an adaptive attention mechanism and dynamic token pruning, which optimizes computational efficiency while maintaining high accuracy. The model is evaluated on five widely used datasets: FER2013, CK+, AffectNet, RAF-DB, and AFEW. It achieves state-of-the-art performance, with accuracies of 98.67% on FER2013, 99.52% on CK+, 99.3% on AffectNet, 96.3% on AFEW, and 98.45% on RAF-DB. An ablation study further validates the contribution of each model component, and comparisons with CNN-based and transformer-based approaches confirm the effectiveness of the model. These findings establish the proposed method as a significant advancement in FER, which offers a scalable and efficient solution for real-world applications.