Facial age estimation by using stacked feature composition and selection

Abstract

In this paper we propose a novel hierarchical feature composition and selection model used in facial age estimation. In recent years, hierarchical architectures have been shown to outperform the flat structure on a variety of visual modeling tasks and has drawn a lot of attention. In our hierarchical architecture, we use biological inspired features as primitive features, then alternatively select and composite newer features. Firstly, we select features in a boosting way and then weightily combine adjacent selected features. The whole process of feature selection and combination is called a boosting layer. We then stack multiple boosting layers into a hierarchical model. In each boosting layer, a number of weak classifiers comprise the selected features, and their combination weights are inversely proportional against the training errors of weak classifiers. In this way, features of a high layer are more descriptive and with higher semantics, while features of a lower layer include more physical details. We expect that this kind of structural features will be more expressive and objective and hence perform better and with higher efficiency in facial age estimation. Our experimental results on two aging face databases MORPH and FG-NET have shown significant reduction on mean absolute error of age estimation compared with other state-of-the-art methods.