Application of Facial Feature Localization Using Constrained Local Models in Template-Based Caricature Synthesis

A three-step method for quasi-frontal face caricature generation is presented in the paper. First a CLM (constrained local models) and CQF (convex quadratic fitting) based face alignment approach is utilized to obtain initial locations of key facial features. Then facial components (eyebrow, eye, nose and mouth) are classified respectively into different categories defined by configuration and appearance in consistent with Chinese-Physiognomy. The mapping relationship between feature attributes and corresponding categories is learnt from a preclassified training set using the decision tree classification algorithm. Finally, separate facial feature cartoon templates are selected according to the classification results and then assembled to form an expressive caricature. Experimental results prove that the presented method is practical and robust for face caricature generation applications.