AI for Fun: Machine Learning Makes Caricature Faces Posted on : Dec 15 - 2018

A team of machine learning developers has created a system for making caricatures of people’s faces.

In a caricature, the artist creates a drawing of a face, with different parts of it greatly enlarged, or exaggerated, in other ways. The pictures or images are usually made to make the person look funny.

Such drawings can be difficult for machines to produce. This is often because the human face is made up of complex shapes with a lot of extremely small details. Generally, good caricatures require skilled artists who can effectively exaggerate faces, while still staying true to life.

The new system for making caricatures is the creation of computer scientists from the City University of Hong Kong and Microsoft Corporation. They presented their findings at a recent computer graphics conference in Tokyo.

Kaidi Cao led the research. He is currently studying computer science at Stanford University in California. He helped develop the system while working as a volunteer at Microsoft. Other members of the team were Lu Yuan of Microsoft and Jing Liao of the City University of Hong Kong.

Cao describes the process as a “photo-to-caricature translation.” He says other machine systems for creating caricatures depend on a series of rules, which are based on how people would draw. But his team’s system uses machine learning to create the caricatures from thousands of examples made by experienced artists.

In a paper explaining his research, Cao said other machine-based systems appear to pay more attention to appearance and style. But his team’s research uses a second machine learning tool to add “geometric exaggeration.”

The new method improves on existing ones “in terms of visual quality and preserving identity,” Cao wrote. In addition, the system gives users more ability to change results in both shape exaggeration and appearance, he added.

The team tested how recognizable its caricatures were to human test subjects in two separate studies.

In the first study, people were shown caricatures and then asked to choose the correct one from a group of images with faces that looked very similar. View More