Feature constraint reinforcement based age estimation

As one of the critical biological characteristics of human age, the face has been widely studied for age prediction, which has broad application prospects in the fields of commerce, security, entertainment, etc. Duo to complicated multi-latent heterogeneous features(e.g. gender) bring valuable messages for the image-based age estimation. A variety of methods utilize heterogeneous information for age estimation. However, heterogeneous features may have uncertain noise, and exploiting them without evaluating the reliability of confidence influence may impact the estimation accuracy. Inspired by the observation that gender has a noticeable impact on face at some particular age stage, this paper proposes a Feature Constraint Reinforcement Network (FCRN) to take advantage of constraint gender influence on the age estimation. The model extracts multi-scale latent heterogeneous features and deduces their confidence of influence upon age estimation methods. Specifically, it gets the gender and age features by classification and regression. Then, the model uses the gender factors extracted from the constraint gender features to reinforce and calculate the influence of different genders on age predictions among different age groups and improve the result of age prediction. Extensive experiments were conducted on the existing public aging datasets. The results show the effectiveness and superiority of the proposed method.