FACE SYNTHESIS USING GENERATIVE ADVERSARIAL NETWORK

Abstract

Generative Adversarial Networks that fall under the class of generative models, aim at taking training examples from the training set and learning the probability distribution that generates those samples. The network is adversarial in the sense that the discriminator tries to maximize the probability of identifying the real data whereas the generator tries to fool the discriminator by producing synthetic data as close as possible to the ground truth value. In recent years, many powerful models using neural network architectures have been introduced that try to learn the discriminative features of the text representations. Also, GANs have been extremely successful in generating realistic images belonging to various categories. Influenced by the success of GANs, researchers thought of applying the GAN model in the human face synthesis task. There exist several attempts for the face synthesis task that try to generate real human faces from the form of input given to them . Unlike the already existing attempts to create human faces, our model tries to apply the concept of text-to-image synthesis [1] GAN in the generation of human faces from the text description stating the attributes o of their respective faces provided as the input. The training of Generator is assisted by the adversary Discriminator (Matching-aware Discriminator model(CNN)) that differentiates the results given by the Generator and the ground truth values. The Generator model would thus learn to generate the human faces that are similar to the ground truth values and thus try to cheat the adversary. The aim is to produce strong results and see the behavior of GAN model in the human face generation task.