Performance Evaluation of Human Facial Expression using Various Classification Methods

Abstract

This study assesses and contrasts the efficacy of raw picture pixels and image vectors as features in face expression classification. The CKPLUS dataset is utilized, and the issue of class imbalance is tackled by data augmentation. The dataset is partitioned into a 70% training set and 30% validation set. The training set consists of 175 images for each class, while the validation set consists of 75 images. The features are displayed using Matplotlib for raw pixels and t-SNE for vector features, then categorized using Random Forest and CNN classifiers. The performance is evaluated by utilizing confusion matrices, accuracy, precision, recall, and F1-score. The findings indicate that the Random Forest algorithm, when combined with vector features, obtains the maximum level of accuracy (99.6190%). Additionally, CNNs using raw pixel features also demonstrate strong performance. The precision, recall, and F1-scores exhibit similarity among the different approaches, with Random Forest (vector feature) and 2D CNN (raw pixels) showing somewhat better performance compared to other methods. These findings suggest that vector features have superior performance when used in conjunction with Random Forest, whereas raw pixel features are more successful when utilized with CNN.