Relationships Self-Learning Based Gender-Aware Age Estimation

In biometrics research, face-appearance based age estimation (AE) becomes an important topic and has attracted a great deal of attention due to its potential applications. To achieve the goal of AE, a variety of methods have been proposed in the literature, among which the cumulative attribute (CA) coding based methods have achieved promising performance by preserving both ordinality and neighbor-similarity of ages. However, the sub-regressors responsible for regressing each of the CA coding elements are learned separately, while all of them are trained on the same dataset, leading to that potential correlation relationships of inter/intra-CA coding are not exploited. To this end, we herein propose a novel correlation learning method to model and utilize such inter/intra-CA relationships for AE, through self-learning from the training data. In addition, we extend the proposed method to perform gender-aware AE by further exploiting the correlations between and within gender groups. Furthermore, we introduce an alternating optimization algorithm for the proposed methods. Extensive experiments are conducted to demonstrate that the proposed methods can significantly improve the accuracy of AE, and more importantly that they can model well both inter/intra CA coding and gender relationships, regardless whether they are related (positive or negative) or not.