Feature Extraction using Deep Learning and Analyses of Curvature on Facial Shapes across Two Races and between Males and Females
Daiki Yamada, Toshinobu Harada, Akira Yamada, Nikhil SHAH, Emily S. Chwa, Sophia G. Allison · Transactions of Japan Society of Kansei Engineering · 2024
In plastic surgery for facial reconstruction and gender conformity, the aspect of appearance of a natural-looking male/ female face is an important factor in the perfection of the surgery. However, there is a problem that the perfection of the postoperative facial shapes after surgery is greatly influenced by the skill of each plastic surgeon. Therefore, it is useful to verify the male/female areas of each patient’s face in order to create an appropriate shape for each patient. In this study, we generated 100 cross-sectional images per person from 3D models of male and female faces, and trained a convolutional neural network (CNN) using gender and race as the classification criteria The trained CNN was then used to visualize the acquired facial features using Grad-CAM and analyze the feature curves. The results revealed that the characteristics of the curves in specific facial regions represent the gender and racial traits.