Machine Learning Generated Synthetic Faces for Use in Facial Aesthetic Research.

Abstract

Importance: A centralized repository of clinically applicable facial images with unrestricted use would facilitate facial aesthetic research. Objective: Using a machine learning neural network, we aim to (1) create a repository of synthetic faces that can be used for facial aesthetic research and (2) analyze synthetic faces according to contemporary aesthetic principles. Design, Setting, and Participants: Synthetic facial images were generated using an open source generative adversarial network. Images were refined and then analyzed using computer vision technology. Interventions: Not applicable. Main Outcomes and Measures: Synthetic facial images were created for use as a facial aesthetic research data set. Results: One thousand synthetic images were generated, and 60 images underwent analysis. Image attributes, including age, gender, image principle axis, facial emotion, and facial landmark points, were attained. Images demonstrated accordance with contemporary aesthetic principles of horizontal thirds and vertical fifths. Images demonstrated excellent correspondence when compared with real human facial photographs. Conclusions and Relevance: We have generated realistic synthetic facial images that have potential as a valuable research tool and demonstrate similarity to real human photographs while adhering to contemporary aesthetic principles.