Existing age estimation methods of performance measurement are mainly based on the training set and testing set of independent identically distributed hypothesis. In order to better conform to the actual scene and better assess the age estimation method of generalization performance， a kind of heterogeneous data sets to evaluate agreement is put forward， i.e. paying more attention to the training set and test set with different distribution and characteristics. In addition， in order to improve the age estimation method based on convolution neural network fitting ability， on the basis of fully considering the adjacent age characteristics， a new theory of loss function analysis is proposed through modeling the age estimation problem as the label distribution study based on Gauss model. Theoretical analysis and experimental results show the effectiveness and robustness of the proposed method.