Double layer multiple task learning for age estimation with insufficient training samples

One of the main difficulty of facial age estimation is the lack of training sample problem. In this paper, we point out that when age estimation is treated as a multiple task learning (MTL) problem, the impact of training sample problem can be relieved. By this idea, we re-formulate the age estimation task using the multi-class score function and develop a double layer multiple task learning (DLMTL) approach. In the subject layer, the personalized age estimation models as well as the global model are used to share knowledge of common aging pattern among different subjects; in the age label layer, the sub-tasks of score function estimation on any specific age label are further modeled to fully exploit the sequential information along the age axis. The proposed DLMTL model can be formulated into a very concise inner product representation, and it is finally solved using the multiple kernel learning (MKL) tool. The experimental results upon the FG-NET and MORPH aging databases verified that our method outperforms many other popular age estimation algorithms especially for the extremely training sample insufficient applications.