Human Age Prediction from Facial Image Using Transfer Learning in Deep Convolutional Neural Networks

Abstract

Automatic age estimation task has attracted attention due to its numerous applications particularly in the field of social media and e-commerce. In this study, a pipeline method of age estimation from human facial images has been investigated where different pre-trained deep convolutional neural networks (DCNNs) are managed through transfer learning. Age may be represented by an integer or floating-point number, but it has some coherence; facial images of few consecutive years of an individual are not so different; even human eye could differentiate a little. Therefore, age estimation of the present study is performed as both regression and classification tasks to show which method preserves more coherence. Only additional layer(s) to the pre-trained DCNNs are reformed for this purpose. Different year groupings (individual, five and ten) are also considered in case of classification. In the proposed method, different DCNN versions of ResNets, Inception and DenseNet are considered on cross-age celebrity dataset (CACD), UTKFace and FGNet datasets. The proposed method is shown to achieve remarkable performance while compared with the existing methods.KeywordsAge predictionTransfer learningPre-trained deep CNNsClassificationRegression