Human Age Estimation by Considering both the Ordinality and Similarity of Ages

The age sequence of human beings exhibits two striking characteristics: ordinal in age values and similar in facial appearance of neighboring ages. Although it has been demonstrated that such ordinality especially the neighboring similarity has positive influence on age estimation, existing approaches have yet not simultaneously taken the two types of information into the estimation. In this paper to conduct age estimation with considering both the ordinality and the neighbor similarity which we call soft-age-contribution (SAC), we take the widely used discriminant method LDA and the least squares regression (LS) as the research baseline, respectively. Firstly, we construct inequality-based large margin ordinal constraints and equality-based ordinal regression constraints and, respectively, incorporate them into LDA and LS to develop their respective ordinal counterpart, coined as OrLDA and OrLS. Next, in order to utilize the SAC information, we formulate two types of membership function to depict such neighboring similarity and embed them into OrLDA and OrLS, yielding soft and ordinal variants of LDA and LS, called SAC-OrLDA and SAC-OrLS in which both the ordinality and the neighboring similarity of ages are considered. Finally, through experiments on benchmark aging datasets, we demonstrate the effectiveness of our strategies in utilizing the two types of information to improving age estimation. In addition, we also quantitatively explore the similarity of neighboring ages, finding that generally about neighboring four years are similar in facial appearance to each other.