A novel hierarchical decomposition model for facial caricature synthesis

This paper presents a hierarchical decomposition model for frontal face caricature generation. The proposed model represents facial features within the framework of a three-layer parsing tree over multiple resolutions. In the face, facial component and texture structure layers of the model, the CLM (Constrained Local Models) based facial feature localization method, the multiple SVM (Support Vector Machine) classifiers fusion based facial component classification algorithm and the NCC (Normalized Cross Correlation) based texture structure analysis approach are utilized to obtain adequate information of the global, partial and detailed facial features respectively. All the facial features analyzed from the above-mentioned coarse-to-fine procedure are associated with corresponding grammatical descriptions defined in a rendering dictionary generalized from the training set that maps each grammatical description to a sub caricature template. According to the associated grammatical descriptions of a face, corresponding sub caricature templates are synthesized and rendered to form an expressive caricature. Experimental results prove that the proposed method is practical and robust for facial caricaturing applications.