Partial Facial Re-imaging Using Generative Adversarial Networks

Existing facial recognition software relies heavily on using neural networks to extract key facial features to accurately classify known individuals. Some of these key features include the shape, size, and distance between an individual’s eyes, nose, and mouth. When these key features cannot be extracted due to facial coverings, existing applications become inaccurate and unreliable. The accuracy and reliability of these technologies are growing concerns as the facial recognition market continues to grow at an exponential rate. In this thesis, we have developed a web-based application service that is able to take in a partially covered face image and generate a new image of what this person could look like without any facial coverings. This service is aimed to be used as an intermediary step between obtaining partially covered face imagery and using facial recognition to accurately classify the individual. This research uses various Generative Adversarial Networks (GAN) to generate facial images of an individual based on preprocessed data extracted from the original image. The web application also allows users the ability to upload data that will be used to build a new GAN model that more accurately represents their needs. This highly scalable service will enable transfer learning and encourage a community of research to be built around this topic