Age Estimation By Multi-Scale Convolutional Network

In the last five years, biologically inspired features (BIF) always held the state-of-the-art results for human age estimation from face images. Recently, researchers mainly put their focuses on the regression step after feature extraction, such as support vector regression (SVR), partial least squares (PLS), canonical correlation analysis (CCA) and so on. In this paper, we apply convolutional neural network (CNN) to the age estimation problem, which leads to a fully learned end-to-end system can estimate age from image pixels directly. Compared with BIF, the proposed method has deeper structure and the parameters are learned instead of hand-crafted. The multi-scale analysis strategy is also introduced from traditional methods to the CNN, which improves the performance significantly. Furthermore, we train an efficient network in a multi-task way which can do age estimation, gender classification and ethnicity classification well simultaneously. The experiments on MORPH Album 2 illustrate the superiorities of the proposed multi-scale CNN over other state-of-the-art methods.