2022 年 21 巻 2 号 p. 111-116
Numerous studies in recent years have focused on the prediction of attractiveness using facial features. Researchers have reported high prediction accuracies using convolutional neural networks (CNNs). In this study, a model was built for predicting the attractiveness of faces using a CNN. In addition, the constructed model is used to visualize the features that underlie the prediction of attractiveness values using the gradient-weighted class activation mapping technique, and the trends are compared with a saliency map applying spectral residual and fine-grained methods. In addition, the relationship between image features that are important for attractiveness and facial features that affect human perception of attractiveness are discussed based on earlier findings in the field of psychology. The results support the psychological finding that sexual dimorphism influences the perception of facial attractiveness, a feature that is important for attractiveness prediction. The approach used here can help us understand the connection between facial features and the perception of attractiveness.