Effectiveness of Data Augmentation in Multi-class Face Recognition

Recent research shows that classification results of using deep learning methods provide better accuracy than those of using machine learning techniques. However, these results are highly dependent on the amount of data used. This study examines the effectiveness of data augmentation in improving the performance of convolutional neural net- work models, namely ResNet-50, MobileNetV3 Small, and MobileNetV3 Large. The performance is measured using the F1-score of the three models of classifying student faces. The experimental results indicate that the F1-score of the ResNet-50 architecture increases from 0.753 to 0.928 when trained using an augmented dataset. The accuracy of the models constructed using the MobileNetV3 Small and Large architectures increased from 0.685 to 0.920 and from 0.660 to 0.930, respectively, when trained using an augmented dataset. The accuracies of ResNet-50, MobileNetV3 Small, and MobileNetV3 Large when evaluated using the testing dataset were 0.865, 0.850, and 0.870, respectively.