Comparative Study of Human Age Estimation Based on Hand-Crafted and Deep Face Features

Abstract

This paper introduces a comparative study of age estimation based on the analysis of facial images. The main contributions are as follows. First, we provide performance evaluation of eight face descriptors which are given by three hand-crafted features as well as by five pre-trained deep Convolutional Neural Networks (CNNs). Second, we show that the use of deep features provided by pre-trained CNNs can transfer the power of the net to new domains and datasets that were not available at the training phase. This leads to an efficient and stable solution to the problem of cross-database by only retraining the regressor instead of the whole network. The experiments are conducted on two public databases: MORPH II and PAL.