Multi-race age estimation based on the combination of multiple classifiers

A considerable amount of research has been conducted on gender and age estimation from facial images over the last few years, and state-of-the-art technology has accomplished a practical accuracy level for a homogeneous race such as Japanese or Korean. However, achieving the same accuracy level across multiple races such as Caucasian, African American, and Hispanic is still highly challenging because of the strong diversity of the growth process of each race. Furthermore, difficulty of gathering training samples uniformly over various races and age brackets makes the problem even more challenging. In this paper, we propose a novel age estimation method that can overcome the above problems. Our method combines a recently proposed machine learning technique called Least-Squares Probabilistic Classifier (LSPC) with neural networks. Through large-scale real-world age estimation experiments, we demonstrate the usefulness of our proposed method.