Face beauty analysis via manifold based semi-supervised learning

Abstract

Beauty has always played an important role in society, implicitly influencing the hu- man interactions of our daily lives and more significant aspects, such as the mate choice or job interviews. And now, with the progress made in deep learning and fea- ture extraction, automatic facial beauty analysis has become an emerging research topic too. However, the subjectivity of beauty still hinders the developement in this area, due to the cost of collecting reliable labeled data, since the beauty score of an individual has to be determined according to various raters. To address this problem, we study the performances of four different semi-supervised manifold based algorithms, which can take advantage of both labeled and unlabeled data in the training phase, and we use them in two different datasets: SCUT-FBP and M 2 B. The learning algorithms are Local and Global Consistency, Flexible Man- ifold Embedding and Kernel Flexible Manifold Embedding. There is an additional algorithm, which, unlike the rest of them, instead of performing classification, ob- tains a non-linear transformation of the data to make the classification easier. All of these algorithms were designed to work on discrete classes, but we perform regres- sion, where labels are real numbers. So the first step, in chapter 2, is to analyse how the algorithms can be adapted to regression and to hypothesize which problems we could be encountering in this process. Secondly, we empirically test them (chapter 3). The best results are obtained with KFME on both datasets, achieving a mean average error of 0.0104 (out of 1) and a Pearson correlation of 0.9782 on SCUT-FBP dataset. With respect to M 2 B dataset, a mean average error of 0.0697 and a Pear- son correlation of 0.7757 are achieved on eastern faces, while a mean average error of 0.0717 and a Pearson correlation of 0.7848 are achieved on western faces. This dissertation ends with a final chapter discussing the results and proposing new topics of study for future work.