A Comparative Study of Hybrid CNN and Vision Transformer Models for Facial Emotion Recognition

Facial emotion recognition is an important image classification problem, popularized by the FER2013 competition. With advancements in deep learning, researchers have shifted from traditional manual algorithms to deep learning models for FER. By implementing deeper and more sophisticated Convolutional Neural Networks (CNNs), highly abstract features can be extracted, leading to significant improvements in Facial Expression Recognition accuracy.In recent years, the Transformer model, initially prominent in natural language processing, has made significant strides in computer vision. This paper investigates the performance of various hybrid models combining CNNs and Vision Transformer (ViT) architectures for FER, with an emphasis on achieving reliable and robust emotion recognition. The experimental results on the FER2013 dataset demonstrate that these hybrid models provide enhanced accuracy and stability across different emotion categories, highlighting the effectiveness of these approaches for practical, real-world applications where consistency is crucial.