Development of Convolutional Neural Network Models to Improve Facial Expression Recognition Accuracy

Abstract

Advancements in information and computer technology, particularly in machine learning, have significantly alleviated human tasks. One of the current primary focuses is facial expression recognition using deep learning methods such as Convolutional Neural Network (CNN). Complex models like CNNs often encounter issues such as gradient vanishing and overfitting. This study aims to enhance the accuracy of CNN models in facial expression recognition by incorporating additional convolutional layers, dropout layers, and optimizing hyperparameters using Grid Search. The research utilizes the FER2013 public dataset sourced from the Kaggle website, trained and evaluated using CNN models, hyperparameter tuning, and downsampling methods. FER2013 comprises thousands of facial images representing various human expressions, with a specific focus on four facial expression categories (angry, happy, neutral, and sad). Through the addition of convolutional and dropout layers, as well as hyperparameter optimization, the developed model demonstrates a significant improvement in accuracy. Findings reveal that the refined CNN model achieves a highest accuracy of 98.89%, with testing accuracy at 89%, precision 78%, recall 78%, and F1-score 78%. This research contributes by enhancing facial expression recognition accuracy through optimized CNN models and providing a framework beneficial for the social-emotional development of children with special needs and aiding in the detection of mental health conditions. Additionally, it identifies avenues for future research, including exploring advanced data augmentation techniques and integrating multimodal information. Furthermore, this study paves the way for applications across diverse fields like human-computer interaction and mental health diagnostics.