Dr. John Femiani is working on GAN-based image compositing

GAN is the abbreviation for Generative adversarial network. This is a piece of machine learning framework in which two neural networks contest with each other. In other words, the computer can be fed information, in this instance faces, and the two networks will work against each other to ‘train’ them to pick out what is versus what isn’t a face.

The hope is that it will be able to take in a plethora of photographs, and use the information to recreate an image that is entirely created, and yet looks realistic to the untrained eye.

Dr. Femiani, an Associate Professor of Computer Science and Software Engineering here at Miami, has been working with these tools in an attempt to create AI that can map your face and allow you to add things into the image that not only look real, but move with you.

Femiani started with a background in art, and thought computer graphics, grew interested in computer science. He got a PhD in computer graphics from Arizona State University.

In a way, he is now trying to merge computer graphics and computer vision in order to create a machine capable of generating images. What makes this different from the typical computer programming is it is not just classification. Instead of simply putting in images and asking the computer to figure out which ones have a face in them, it is being told to make a face in its entirety.

There are still some things to work through, specifically things that can confuse the computer, such as a person wearing jewelry, which can throw off the AI’s typical perception of what a face looks like.

It took about two weeks for Dr. Femiani and his team to ‘train’ their computer to recognize faces. Instead of simply creating faces, however, they want to use refinement (or transfer learning) to add things into the image that look realistic.

The difficulty often comes about when trying to render hair, wrinkles, moles, piercings, etc. that still look realistic and clear, while having the proper lighting and other things that will blend it seamlessly into the photo.

What Femiani and his team have been focusing on recently is the ability to put any hairstyle over someone's face.

The hope for the future is that Dr. Femiani and his team will be able to correct any issues with the software, such as making it work for all skin types and hair textures, as well as using faces from different angles.

You can watch a youtube video here for more about the research. You can also read the research paper here.

Written by Kayleigh Schauseil, CEC Reporter