Deep Rank Learning for Facial Attractiveness

Predicting facial attractiveness automatically is a rising problem and has received more attention from the computer vision community. Previous works have posed facial attractiveness as a ranking problem. However, extracting features and learning a ranking model are often separated and optimized independently in previous works. In this paper, we propose two End-to-End learning to rank approaches using deep Convolutional Neural Network for fully automatic facial beauty assessment, which are RankSVM CNN and RankNet CNN. To the best of our knowledge, this is the first work to address learning to rank problem for facial beauty by CNN. We proved both the RankNet CNN and the RankSVM CNN are effective and excellent for facial beauty assessment. With a small amount of data, our approaches perform better than any other shallow models we have known. Finally, we have achieved a high prediction correlation, which is 0.469 on the HOTorNOT dataset and 0.815 on the SCUT-FBP dataset.