Corneal Localization for Discerning Faces using Advanced Machine Learning Algorithms of GAN

The field of Deep Learning (DL) and Machine Learning (ML) has evolved over the years, these models generally consist of biometric recognition systems which need an extensive dataset, high-performance computing systems, and a time- consuming process approach. With existence of modern frameworks such as Generative Adversarial Networks, numerous features in biometric systems which are employed for data security can be spoofed into, as a result it creates unexplored vulnerability. These models are extremely sophisticated and are now able to manufacture artificial human features which are difficult to differentiate from the authentic ones. The work in this paper helps to recognize and explore various GAN-synthesized corneal features with the use of style generative adversarial network 2 (StyleGAN-2). The model presented extract features of eyes from the facial images passed into it, Specifically, the cornea which consists of specular highlights of an individual. The training method used for the model is employed by generating multiple unique data instances, which synthesize enormous high quality dataset containing features resembling the training data. Thus, simplifying the process of training the model in an efficient manner by making use of progressive generators in StyleGAN-2 with CNN algorithms. Generated images and real images are further provided to the discriminator for discerning the model-synthesized eyes from real eyes. The generation of synthesized images, to the extraction of corneal specular highlights was experimented using three different processing methods: Central Processing Unit (CPU), Graphical Processing Unit (GPU) and a Tensor Processing Unit (TPU). Comparative analysis on these three methods was performed based on performance and time taken when tested with varying quantity of datasets and diverse environment.