Semi-supervised auto-encoder for facial attributes recognition

Abstract

The particularity of our faces encourages many researchers to exploit their features in different domains such as user identification, behaviour analysis, computer technology, security, and psychology. In this paper, we present a method for facial attributes analysis. The work addressed to analyse facial images and extract features in the purpose to recognize demographic attributes: age, gender, and ethnicity (AGE). In this work, we exploited the robustness of deep learning (DL) using an updating version of autoencoders called the deep sparse autoencoder (DSAE). In this work we used a new architecture of DSAE by adding the supervision to the classic model and we control the overfitting problem by regularizing the model. The pass from DSAE to the semi-supervised autoencoder (DSSAE) facilitates the supervision process and achieves an excellent performance to extract features. In this work we focused to estimate AGE jointly. The experiment results show that DSSAE is created to recognize facial features with high precision. The whole system achieves good performance and important rates in AGE using the MORPH II database.