MT-ResNet: A Multi-Task Deep Network for Facial Attractiveness Prediction

Abstract

Facial attractiveness prediction (FAP) is an intriguing and challenging problem that draws attention of researchers in recent years. Unlike other objective computer vision topics such as face detection, FAP also involves deep facial feature extraction and attractiveness pattern recognition which is relatively subjective. The work of FAP requires both mass collection of people's appreciations of beauty and the learning, replication of people's aesthetic standards by the model. Work regarding FAP in the early stage focuses on representing facial features using machine learning algorithms. In recent years, neutral networks, especially convolutional neural networks show its great performance in related areas. In this paper, a multi-task FAP model, MT-ResNet is proposed which could automatically predict the facial attractiveness score and the gender given a portrait. The results are compared with other existing models, which shows MT-ResNet's efficiency and high-accuracy among similar works.