Facial age estimation based on asymmetrical label distribution.

Facial age estimation is promising for a wide range of applications including security access control, biometrics, and human–computer interaction. Current label distribution learning-based age estimation methods usually assume a prior probability distribution to handle the correlation among adjacent ages, and Gaussian distribution is the most widely used one. However, the Gaussian distribution is symmetrical, it does not conform to the diversified characteristic of age processing. In this paper, we propose an age encoding method called ALD (asymmetrical label distribution) by fusing the apparent bimodal distributions and chronological single distribution so that much more robust age label distribution can be found. Then, a lightweight multi-task learning network is designed to perform label distribution learning and regression learning for both global estimation and stagewise estimation. Unlike traditional cascade networks of label distribution learning, the parallel structure of multi-task is able to reduce the error propagation from predictive distribution to regression. Extensive experimental analyses on four benchmark datasets demonstrate the superior performance of the proposed method. [ABSTRACT FROM AUTHOR]