Performance Analysis of GAN Architecture for Effective Facial Expression Synthesis

Abstract

Facial expressions are an important means of communication among human beings, as they convey different meanings in a variety of contexts. All human facial expressions, whether voluntary or involuntary, are formed as a result of movement of different facial muscles. Despite their variety and complexity, certain expressions are universally recognized as representing specific emotions - for instance, raised eyebrows in combination with an open mouth are associated with surprise, whereas a smiling face is generally interpreted as happy. Deep learning-based implementations of expression synthesis have demonstrated their ability to preserve essential features of input images, which is desirable. However, one limitation of using deep learning networks is that their dependence on data distribution and the quality of images used for training purposes. The variation in performance can be studied by changing the optimizer and loss functions, and their effectiveness is analysed based on the quality of output images obtained.