Improvement of Facial Beauty Prediction Using Artificial Human Faces Generated by Generative Adversarial Network

Human beauty evaluation is a particularly difficult task. This task can be solved using deep learning methods. We propose a new method for determining the attractiveness of a face by using the generation of synthetic data. Our approach uses the generative adversarial network (GAN) to generate an artificial face and then predict the facial beauty of the generated face to improve facial beauty predictions. A study of images with different brightness and contrast showed that the methods using the convolutional neural network (CNN) model have fewer errors than compared to the multilayer perceptron (MLP) model that uses the method. The MLP model only responds to geometric facial proportions, whereas the CNN model additionally responds to changes in face color. Using the synthetic face instead of the real face improves the determination of accuracy of the facial attractiveness. The ability to appreciate facial beauty also opens the way for facial beauty modifications in a latent space. Further research could improve facial normalization in the latent space to improve the accuracy of facial beauty determination.