Optimasi Feature Selection Dalam Sistem Pengenalan Emosi Wajah Menggunakan CNN dan GNN

Abstract

Facial Emotion Recognition (FER) is a field of computer vision that aims to identify human emotions based on facial expressions. Convolutional Neural Networks (CNNs) are widely used for automatic visual feature extraction; however, they have limitations in modeling spatial relationships among facial components. Therefore, this study implements a Graph Neural Network (GNN) based on the Graph Attention Network (GAT) architecture and applies Principal Component Analysis (PCA) as a feature selection method to improve feature representation quality and computational efficiency. The study utilizes the FER-2013 dataset, which consists of seven emotion classes: angry, disgust, fear, happy, sad, surprise, and neutral. The research stages include data preprocessing, feature extraction using MobileNetV2, dimensionality reduction using PCA, graph construction, model training, and performance evaluation using accuracy, precision, recall, and F1-score metrics. Four experimental scenarios were evaluated: CNN, CNN+PCA, GNN, and GNN+PCA. The results show that the CNN model achieved an accuracy of 62%, CNN+PCA achieved 61.95%, GNN achieved 70.71%, and GNN+PCA achieved 74.32%. Furthermore, the GNN+PCA model obtained a precision of 75.57%, recall of 74.32%, and F1-score of 74.64%, representing the best performance among all evaluated models. The findings indicate that PCA does not significantly improve CNN performance but substantially enhances GNN performance by reducing feature redundancy and optimizing graph-based feature representation. In conclusion, the combination of GNN and PCA proved to be more effective than CNN and CNN+PCA in recognizing facial emotions on the FER-2013 dataset. Therefore, this approach has strong potential as a more accurate deep learning-based facial emotion recognition system.

GS Cited Analysis