A Study on Visualizing Facial Attractiveness Factors Using Deep Learning Methods

抄録

In recent years, there has been significant research on the use of deep learning to predict facial attractiveness and beauty. Such studies are expected to result in several applications. To improve the prediction accuracy, it is necessary to investigate which facial features are predictors. The purpose of this study was to identify features that are important for facial attractiveness prediction models using two visualization methods: Gradient-weighted Class Activation Mapping (Grad-CAM) and Grad-CAM++. For male images, Grad-CAM showed activity around the eyebrows, whereas Grad-CAM++ also showed activity in the eyes, eyebrows, and skin regions. For female images, Grad-CAM showed activity around the eyes and forehead, and Grad-CAM++ showed activity around the eyes and forehead in some images. These results are consistent with psychological findings, and such methods may facilitate the understanding of facial attractiveness factors.