Extraction of Features Important for Facial Attractiveness Using Gradient-Weighted Class Activation Mapping and Guided Gradient-Weighted Class Activation Mapping

Abstract

Numerous studies have been conducted to determine the factors that contribute to facial attractiveness. In recent years, interest in using deep learning to predict facial attractiveness and extract features that are important for such a prediction has increased. In this study, the face attractiveness prediction model visualizes features that are important for prediction via two methods, i.e., gradient-weighted class activation mapping (Grad-CAM) and guided Grad-CAM, and then the results are compared. The results show that Grad-CAM visualizes primarily the feature space that is important for attractiveness prediction, whereas guided Grad-CAM visualizes more detailed information such as contours. These methods may facilitate the understanding of facial attractiveness factors.