Age estimation with expression changes using multiple aging subspaces

Image-based human age estimation has become one of the interesting but challenging problems in computer vision and biometrics. It is even harder when the faces have different expressions. In this paper, we propose a weighted random subspace method to solve the relatively new problem: cross-expression age estimation. The proposed method does not depend on the learning of correlation between different expressions, and thus could work in the situation when the expression-correlation does not exist in the training data. We also explore the use of data from multiple datasets to further improve the estimation performance. Experiments on two aging datasets with explicit expression changes demonstrate that the proposed approach gives superior performance over the state-of-the-art method.