Automatic Age Estimation of Persons with Dark Skin Tone Using Deep Learning Approach

Abstract

Age estimation has become very important for many organizations and human endeavour. Traditional machine learning methods have previously been deployed in previous times by many researchers to solve this problem automatically. However, the use of deep learning methods in recent times has shown superior performance in artificial intelligence tasks. This study employs the deep learning method gleaned from the ResNet50 convolutional neural network (CNN) to solve this problem; but on persons with dark skin tone, because pre-existing automatic age estimation models were trained on datasets with a limited population of persons with darker skin tones, as recent studies have shown that the aging features of dark skin tone persons cannot be learnt from persons with light skin pigmentations (white skin tone persons). A combination of persons with dark skin tones from the UTKFace, APPA-REAL and BlackFaces datasets was used to train the CNN. At the end of the experiment, the proposed approach attained a mean absolute error of 5.21 years on the validation set, and showed good performance on age estimation of dark skin tone persons.