Facial Expression Recognition Through Deep Learned Features

This paper investigates improving facial expression recognition (FER) using Convolutional Neural Networks (CNNs) by fine-tuning parameters. The research objectives are to achieve high accuracy, compare different models, and identify the most suitable model for FER. Deep learning methods, especially CNNs, have shown promise in addressing FER challenges. We explore several models on the FER2013 dataset and propose a CNN model with ReLU activation functions, Softmax output layers, Adam and RMSprop optimization, achieving high accuracy with fewer epochs. We analyze strengths and weaknesses of other models, providing insights for future research. The study aims to advance FER systems for enhanced human-computer interaction and emotional analysis in various applications. Results show that the CNN-MLP and VGG-16 models perform moderately, while the CNN-XAI model achieves around 57% accuracy. While addressing overfitting, the proposed CNN model stands out for its accuracy. Future work includes exploring different architectures, hyperparameters, and incorporating alternative evaluation metrics like ROC and AUC for further improvements.