Mixture of deep networks for facial age estimation

Abstract

In this paper, our objective is to simultaneously explore the learning of ordinal relationships among age labels and address the challenge of heterogeneous data resulting from the non-stationary aging process through an advanced mixture model of deep networks. Drawing upon the pivotal insight that the non-stationary aging process can be decomposed into a series of stationary subprocesses, we employ a divide-and-conquer strategy. This involves initially partitioning the age spectrum into multiple groups and subsequently training a specialized deep network, referred to as an “expert”, for each distinct group. These experts are not functionally independent; instead, they are interconnected through specialized model designs and a joint training mechanism that consolidates them into a unified system. As a result, the learning of ordinal relationships is consistently maintained by solving the age-related tasks across the entire age label set. The final age estimation is accomplished through a hierarchical classification approach, leveraging the collective outputs from all the experts. Extensive experiments involving several well-known datasets for age estimation have demonstrated the superior performance of our proposed model over several existing state-of-the-art methods.