Facial Emotion Recognition Using Masked-Augmented FERPlus

The Masked-Augmented FERPlus dataset seeks to resolve the issue of dataset imbalance in emotion detection, particularly when faces are obstructed by objects such as masks. The objective is to determine the appropriate oversample size using the classification complexity metric to address the dataset imbalance in the original FERPlus dataset. The dataset was used to train various machine learning models, including Convolutional Neural Networks (CNN), VGG16, ResNet, and InceptionV3, and their accuracy was measured to determine the most effective model. Although the oversampled dataset demonstrated superior performance with the N1 metric calculation, the model's accuracy on the original dataset, which was 0.752, slightly surpassed its performance on the masked augmented dataset, which achieved an accuracy of 0.747. Despite the simple augmentation technique implemented in this study, the results indicated that it effectively reduced model overfitting while training.