Facial Age Estimation and Life Stage Classification Using Fine-Tuned ResNet-101

Abstract

Automatic age estimation from facial images is a challenging problem due to the complex nature of the aging process and unconstrained acquisition conditions like lighting and pose. In this paper, present a method based on a fine-tuned ResNet-101 convolutional neural network that simultaneously solves a continuous age regression task and a four-class life stage classification (child, youth, adult, senior). Detect and align faces using MTCNN and extract features from two streams, the first being the global face appearance and the second a periocular region centered around the eyes automatically localized by MTCNN. Both streams are projected to a common latent 256-dimensional space, concatenated into a 512-dimensional vector, and used to predict age and life-stage labels with a weighted combination of L1 and cross-entropy losses. Evaluate our method on the UTKFace data set containing 23,708 images. On the hold-out set of 4,738 images, our method achieved a mean absolute error (MAE) of 4.15 years, an overall accuracy of 90.82% and a macro F1-score of 0.895. The F1-scores for the child, adult, and senior classes were between 0.92% and 0.95%. was obtained for the youth class due to confusion with the adult class. Our results demonstrate that using both global and periocular feature streams is a viable approach to solve age-aware tasks such as biometrics and security verification.