Analysis of Face Data Augmentation in Various Poses for Face Recognition Model

When constructing a face recognition model, many face images for each person are required for the model to perform well. Face data augmentation is a popular technique to create variations of face poses via angular head movements, e.g., yaw, pitch, and roll. This study used convolutional neural networks, ResNet-50, MobileNetV3-Large, and SeResNet18 models to evaluate the impact of face data augmentation on the performance of the models. The training was performed using the FaceScrub dataset to evaluate the performance of the face data augmentation approach using a batch size of 20, a learning rate of 0.001, and 150 epochs. The results demonstrate that the SeResNet18 model obtained the highest accuracy (0.8719) with a loss of 1.5106. Before increasing the amount of face data through various poses, the accuracy of the SeResNet18 model was only 0.8108, with a loss of 1.5513. We also discovered that the original data combined with all angular head movements (yaw, pitch, and roll) produced the best result for all models.