Deep learning approaches in face analysis

Abstract

Although face analysis algorithms have changed over the decades, almost in every face-related algorithm it is still usually the case that the order of the problem solving algorithm is the same. The analysis task is relatively easy on frontal and clear faces. However, when it comes to in-the-wild objects with spontaneous expressions, it is a challenging issue due to the changes in illumination, pose variation, expression intensity, subtle deformations, occlusion, etc. The recent success in deep neural networks makes it inevitable to ignore the technique in face analysis to automatically learn the discriminative representations of the face. In a deep network, the input data is run through several hidden layers that decompose the features of the input. The features are then classified using a function to retrieve the class probabilities to predict the output class. This chapter will investigate the deep learning approaches used to detect and pre-process the face, estimate its attributes, classify the expression and recognize the face. © 2022 Elsevier B.V., All rights reserved.