An Ensemble CNN2ELM for Age Estimation

Abstract

Age estimation is a challenging task, because it can be easily affected by gender, race, and other intrinsic and extrinsic attributes. At the same time, performing age estimation for a narrow age range may lead to better results. In this paper, to achieve robust age estimation, an ensemble structure referred to as CNN2ELM, which includes convolutional neural network (CNN) and extreme learning machine (ELM), is proposed for age estimation. The three-level system includes feature extraction and fusion, age grouping via an ELM classifier, and age estimation via an ELM regressor. Age-Net, Gender-Net, and Race-Net are trained using different targets, such as age class, gender class, and race class, respectively, and the three networks are used to extract features corresponding to age, gender, and race from the same image of a person during validation and test stages. Features related to the age property are enhanced by fusing these of race and gender properties. Then, to achieve a narrow age range, the ELM classifies the fusion results into one of the age groups. Afterward, an age decision is made using an ELM regressor. Our network is pretrained on an ImageNet database and then fine-tuned on the IMDB-WIKI database. The recently released Adience benchmark, ChaLearn Looking at People 2016 (LAP-2016), and MORPH-II are used to verify the performance of “Race-Net + Age-Net + Gender-Net + ELM classifier + ELM regressor (RAGN).” RAGN outperforms the existing state-of-the-art age estimation methods. The mean absolute error of the age estimation of RAGN for MORPH-II is determined to be 2.61 years; the accuracy of the age estimation for the Adience benchmark is 0.6649; and the normal score (ϵ) for the sequestered test set of the LAP-2016 data set is 0.3679.