Transfer Learning Adaptive Facial Attractiveness Assessment

Abstract

Recent advances in Computer Vision and Artificial Intelligence have brought the opportunity to automate facial attractiveness evaluation. A range of studies have been addressed to the task and have achieved reasonable prediction accuracy. However, most of these methods work well only on photos with restrictions on expression, posture, illumination, but not on real-world face photos. This work is aimed to improve the attractiveness assessment state-of-the-art in both cases. To this end, an approach that employs transfer learning methodology as well as shallow machine learning was proposed for highly accurate facial attractiveness prediction. Specifically, a Convolutional Neural Network (CNN), Facenet, originally designed and pre-trained for the face recognition task is utilized. High-level facial features were extracted by using the network and then fed into Support Vector Regression in order to predict facial attractiveness. Extensive experiments conducted on widely used facial beauty datasets Gray and SCUT-FBP5500 demonstrated that the proposed method outperformed other attractiveness prediction approaches. The experimental results also confirmed the effectiveness of the method in both constrained and unconstrained environment.