Age estimation from face image using Hybrid Representation for Deep Learning

Abstract

New technique of Deep learning gave Face-age estimation higher results-accuracy. Proposed estimation technique of this paper presents a novel deep learning network for age estimation called Hybrid Representation Architecture (HRA). Where local, global and global-local branches are contained and jointly optimized for capturing multi-type features combined by complementary information. In each branch, the sub-network is designed for to extract features from independent region by employing a separate loss, where recurrent fusion is used for exploring correlations among them. We consider that different fashions of pose would cause misalignment in regions variation, thus an Aligned Region Pooling (ARP) operation is designed to extract aligned-region features. To compensate the demand of large-scale image datasets, a private image dataset regarding-age progression (named Web-FaceAge) is adopted with more than 120K face images, which are captured under different scenes providing wide range of ages. Benchmarking experiments conducted on five image datasets, such as MORPH, CACD, Chalearn LAP 2015, FG-NET and Web-FaceAge, show the significant precedence of proposed estimation technique against other state-of-the-art techniques.