In this paper, the face image is divided into two separated parts: aging effect feature part and identity related feature part. The identity dictionary and the age dictionary are introduced to encode the two feature parts into two separated feature spaces. To make sure the learned dictionaries are discriminative for different classes, the reconstruction error and label matrices constraints are added in the training. Face features can be encoded into identity and age space with the learned identity and age dictionaries. The identity space can be used for further classification. Extensive experiments are conducted on the MORPH and FGNET dataset, illustrating a great improvement over the state-of-the-arts.