This work presents a machine learning approach for facial image-based age estimation. The idea is to first extract age relevant texture and shape features from a set of images and use them in combination with the subject’s age to learn a model of the human aging process. The classification is done in two steps. At first a classification between youths and adults is done. In the second step the exact age is estimated by a more specific classifier based on the result of the first step. Extensive experiments on the FG-NET aging database are conducted using the leave one person out evaluation scheme. Modifications regarding the feature extraction and taking a soft decision in the first step of the classification are found to improve the performance, leading to a mean absolute error of 4.77 years, which is the lowest mean absolute error reported on the FG-NET aging database.