Face Expression Neutralisation Using Improved Variational Autoencoder

There is a high demand for realistic facial expressions in modern computer graphics and multimedia research. Unfortunately, synthesizing face expressions takes time, effort, and hard labour since high naturalism of face expressions is required. Despite recent advances in synthesizing realistic facial images, current generative models conflict to catch more involved image types, potentially due to their latent space simplified architecture. Furthermore, the union of a Generative Adversarial Network (GAN) with a Variational Autoencoder (VAE) by using usually the Adam optimizer has recently been studied and investigated in cutting-edge research. Hence, at tiem of learning the generator takes a long time and then overfit for a particular time instance. So, in this work, we will suggest integrating the VAE to GAN’s architecture by using the Ranger optimizer into an unsupervised generative model that at once learns to encode, generate and compare dataset samples to bring out various real neutral faces in a reduced and optimal time. Next, we proposed the probability distributions called Kullback-Leibler Divergence (KLD) as an objective weighting scheme that helps us to measure just how much information we lose when we choose an approximation during training. Also, the Mean Square Error (MSE), the most commonly used loss function for regression is used as a metric of evaluation for the experiment tests.