CARTOONNET: Caricature Recognition of Public Figures

Abstract

Recognizing faces in the cartoon domain is a challenging problem since the facial features of cartoon caricatures of the same class vary a lot from each other. The aim of this project is to develop a system for recognizing cartoon caricatures of public figures. The proposed approach is based on the Deep Convolutional Neural Networks (DCNN) for extracting representations. The model is trained on both real and cartoon domain representations of a given public figure, in order to compensate the variations in the same class. The IIIT-CFW (Mishra et al., European conference on computer vision, 2016) [1] dataset, which includes caricatures of public figures, is used for the experiments. It is seen from these experiments that improving the performance of the model can be achieved when it is trained on representations from both real and cartoon images of the given public figure. For a total of 86 different classes, an overall accuracy of 79.65% is achieved with this model.