Facial Emotion Recognition: A Comparative Study of Pre-trained Deep Learning Models

This paper presents a comparative study of facial emotion detection methodologies, highlighting a robust approach that encompasses key stages such as face detection, feature extraction, and emotion classification. Central to this approach are convolutional neural networks (CNNs), specifically leveraging architectures of pre-trained models like VGG16 and DenseNet121. These CNNs are adept at automatically learning discriminative features from facial images, enabling the system to accurately discern between various emotional states and assign corresponding labels. Our findings revealed that VGG16 and DenseNet121 emerged as the top performers among the models examined, achieving an impressive accuracy score of 86%, surpassing the ensemble model. In the evaluation phase, the system demonstrates impressive performance metrics, with VGG16 and DenseNet121 consistently outperforming the random ensemble model’s with a 0.86% accuracy. The efficacy of deep learning architectures in capturing nuanced facial expressions and facilitating accurate emotion recognition. The study does, however, admit a number of difficulties with facial emotion recognition. Facial expression variability among people, groups, and cultures presents a major challenge, requiring strong training and generalization of the model. Moreover, any biases or inadequate representation of specific demographics in the FER-2013 dataset could affect the system’s effectiveness and generalizability.Moreover, the computational resources required for training and deploying deep learning models represent another hurdle. The complexity of CNN architectures and the computational intensity of training processes may impose practical constraints, especially in resource-constrained environments.