A Local Feature Oriented Ensemble Method for Caricature Recognition

Caricature is a popular art form. Caricature artists usually exaggerate facial features to highlight the characteristics of people. Although it is not difficult for people to recognize caricature faces even if the face structures are far away from real faces, traditional face recognition methods do not perform well in matching caricature faces with real faces. Based on the Webcaricature database and Residual Neural Network (ResNet), an ensemble method that emphasizes local features of faces is proposed in this paper. Local features are trained separately to form models that evaluate the distances of different features. Then all local feature models are weighted and combined with a full-face model. In order to stabilize the average performance, the weight of each local feature must be controlled within an appropriate range. Through many experiments, the ensemble model that integrates eyes, mouth, and full-face model with adapted weights demonstrated better performance, and the overall accuracy rate of recognition indicated considerable improvement. This paper proposed a promising ensemble method to deal with the current difficulties in caricature recognition.