TY - JOUR
T1 - Age estimation via attribute-region association
AU - Chen, Yiliang
AU - He, Shengfeng
AU - Tan, Zichang
AU - Han, Chu
AU - Han, Guoqiang
AU - Qin, Jing
N1 - Funding Information:
This project is supported by the National Natural Science Foundation of China (No. 61472145, No. 61972162, and No. 61702194), the Innovation and Technology Fund of Hong Kong (Project No. ITS/319/17), the Special Fund of Science and Technology Research and Development on Application From Guangdong Province (SF-STRDA-GD) (No. 2016B010127003), the Guangzhou Key Industrial Technology Research fund (No. 201802010036), the Guangdong Natural Science Foundation (No. 2017A030312008), and the CCF-Tencent Openfund.
Publisher Copyright:
© 2019 Elsevier B.V.
PY - 2019/11/20
Y1 - 2019/11/20
N2 - Human age has been treated as an important biometric trait in many practical applications. In this paper, we propose an Attribute-Region Association Network (ARAN) to tackle the challenging age estimation problem. Instead of performing prediction from a global perspective, we delve into the relationship between face attributes and regions. First, the proposed network is guided by the auxiliary demographic information, as different demographic information (e.g., gender and ethnicity) intrinsically correlates to human age. Second, different face components are separately handled and then involved in the proposed ensemble network, as these components vary differently along with human age. To explore both global and local information, the proposed network consists of several sub-network, each of them takes the global face and a face sub-region as input. Each sub-network leverages the intrinsic correlation between different face attributes (i.e., age, gender, and ethnicity), and it is trained in a multi-task manner. These attribute-region sub-networks are associated to yield the final predictions. To properly train and coordinate such a complex network, a new hierarchical-scheduling training method is proposed to balance the learning complexity in the multi-task learning. In this way, the performance of the most difficult task (i.e., age estimation) can be significantly improved. Extensive experiments on the MORPH Album II and FG-NET show that the proposed method outperforms the state-of-the-art age estimation methods by a significant margin. In particular, for the challenging age estimation, the Mean Absolute Errors (MAE) are decreased to 2.51 years compared to the state-of-the-arts on the MORPH Album II dataset.
AB - Human age has been treated as an important biometric trait in many practical applications. In this paper, we propose an Attribute-Region Association Network (ARAN) to tackle the challenging age estimation problem. Instead of performing prediction from a global perspective, we delve into the relationship between face attributes and regions. First, the proposed network is guided by the auxiliary demographic information, as different demographic information (e.g., gender and ethnicity) intrinsically correlates to human age. Second, different face components are separately handled and then involved in the proposed ensemble network, as these components vary differently along with human age. To explore both global and local information, the proposed network consists of several sub-network, each of them takes the global face and a face sub-region as input. Each sub-network leverages the intrinsic correlation between different face attributes (i.e., age, gender, and ethnicity), and it is trained in a multi-task manner. These attribute-region sub-networks are associated to yield the final predictions. To properly train and coordinate such a complex network, a new hierarchical-scheduling training method is proposed to balance the learning complexity in the multi-task learning. In this way, the performance of the most difficult task (i.e., age estimation) can be significantly improved. Extensive experiments on the MORPH Album II and FG-NET show that the proposed method outperforms the state-of-the-art age estimation methods by a significant margin. In particular, for the challenging age estimation, the Mean Absolute Errors (MAE) are decreased to 2.51 years compared to the state-of-the-arts on the MORPH Album II dataset.
KW - Age estimation
KW - Attribute-region association
KW - Multi-task learning
UR - http://www.scopus.com/inward/record.url?scp=85071322540&partnerID=8YFLogxK
U2 - 10.1016/j.neucom.2019.08.034
DO - 10.1016/j.neucom.2019.08.034
M3 - Journal article
AN - SCOPUS:85071322540
SN - 0925-2312
VL - 367
SP - 346
EP - 356
JO - Neurocomputing
JF - Neurocomputing
ER -