Abstract

We present a system to automatically create a caricature drawing froma frontal face photograph. The system learns how to exaggerate features based on training examples from a particular caricaturist. The input to the system is a frontal face image along with its vector based line-drawing. The system creates a face shape model of the given face with a minimum of user interaction. The face shape is then converted to a set of semantic face features. A Kernel Regression (KR) algorithm determines how much each feature should be exaggerated based on the knowledge learned from a particular artist. The exaggerated features are then projected into an exaggerated face shape using a maximum likelihood estimation (MLE) algorithm. This MLE algorithm also maintains face constraints to generate reasonable caricatures. Finally, the exaggerated face shape together with the unexaggerated face shape are used to morph the input line-drawing to create a caricature. A user interface is provided to indicate the extent of overall exaggeration. We have trained the system with two separate artists and show results of passing several photographs and line drawings through the system.